Skip to content

GitLab

"tests/models/test_hf_steered.py" did not exist on "5e0bc289dfcfe92cd3cd93836c0c863a99f4d566"
Commit a380ace2 authored by chenzk's avatar chenzk
Browse files

v1.0

parents
Show whitespace changes
Inline Side-by-side
Showing with 8679 additions and 0 deletions
build/lib/build/lib/build/lib/unsloth/models/mapper.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
__all__ = [
"INT_TO_FLOAT_MAPPER",
"FLOAT_TO_INT_MAPPER",
]
__INT_TO_FLOAT_MAPPER = \
{
"unsloth/mistral-7b-bnb-4bit" : (
"unsloth/mistral-7b",
"mistralai/Mistral-7B-v0.1",
),
"unsloth/llama-2-7b-bnb-4bit" : (
"unsloth/llama-2-7b",
"meta-llama/Llama-2-7b-hf",
),
"unsloth/llama-2-13b-bnb-4bit" : (
"unsloth/llama-2-13b",
"meta-llama/Llama-2-13b-hf",
),
"unsloth/codellama-34b-bnb-4bit" : (
"codellama/CodeLlama-34b-hf",
),
"unsloth/zephyr-sft-bnb-4bit" : (
"unsloth/zephyr-sft",
"HuggingFaceH4/mistral-7b-sft-beta",
),
"unsloth/tinyllama-bnb-4bit" : (
"unsloth/tinyllama",
"TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
),
"unsloth/tinyllama-chat-bnb-4bit" : (
"unsloth/tinyllama-chat",
"TinyLlama/TinyLlama-1.1B-Chat-v1.0",
),
"unsloth/mistral-7b-instruct-v0.1-bnb-4bit" : (
"unsloth/mistral-7b-instruct-v0.1",
"mistralai/Mistral-7B-Instruct-v0.1",
),
"unsloth/mistral-7b-instruct-v0.2-bnb-4bit" : (
"unsloth/mistral-7b-instruct-v0.2",
"mistralai/Mistral-7B-Instruct-v0.2",
),
"unsloth/llama-2-7b-chat-bnb-4bit" : (
"unsloth/llama-2-7b-chat",
"meta-llama/Llama-2-7b-chat-hf",
),
"unsloth/llama-2-7b-chat-bnb-4bit" : (
"unsloth/llama-2-7b-chat",
"meta-llama/Llama-2-7b-chat-hf",
),
"unsloth/codellama-7b-bnb-4bit" : (
"unsloth/codellama-7b",
"codellama/CodeLlama-7b-hf",
),
"unsloth/codellama-13b-bnb-4bit" : (
"codellama/CodeLlama-13b-hf",
),
"unsloth/yi-6b-bnb-4bit" : (
"unsloth/yi-6b",
"01-ai/Yi-6B",
),
"unsloth/solar-10.7b-bnb-4bit" : (
"upstage/SOLAR-10.7B-v1.0",
),
"unsloth/gemma-7b-bnb-4bit" : (
"unsloth/gemma-7b",
"google/gemma-7b",
),
"unsloth/gemma-2b-bnb-4bit" : (
"unsloth/gemma-2b",
"google/gemma-2b",
),
"unsloth/gemma-7b-it-bnb-4bit" : (
"unsloth/gemma-7b-it",
"google/gemma-7b-it",
),
"unsloth/gemma-2b-bnb-4bit" : (
"unsloth/gemma-2b-it",
"google/gemma-2b-it",
),
"unsloth/mistral-7b-v0.2-bnb-4bit" : (
"unsloth/mistral-7b-v0.2",
"alpindale/Mistral-7B-v0.2-hf",
),
"unsloth/gemma-1.1-2b-it-bnb-4bit" : (
"unsloth/gemma-1.1-2b-it",
"google/gemma-1.1-2b-it",
),
"unsloth/gemma-1.1-7b-it-bnb-4bit" : (
"unsloth/gemma-1.1-7b-it",
"google/gemma-1.1-7b-it",
),
"unsloth/Starling-LM-7B-beta-bnb-4bit" : (
"unsloth/Starling-LM-7B-beta",
"Nexusflow/Starling-LM-7B-beta",
),
"unsloth/Hermes-2-Pro-Mistral-7B-bnb-4bit" : (
"unsloth/Hermes-2-Pro-Mistral-7B",
"NousResearch/Hermes-2-Pro-Mistral-7B",
),
"unsloth/OpenHermes-2.5-Mistral-7B-bnb-4bit" : (
"unsloth/OpenHermes-2.5-Mistral-7B",
"teknium/OpenHermes-2.5-Mistral-7B",
),
"unsloth/codegemma-2b-bnb-4bit" : (
"unsloth/codegemma-2b",
"google/codegemma-2b",
),
"unsloth/codegemma-7b-bnb-4bit" : (
"unsloth/codegemma-7b",
"google/codegemma-7b",
),
"unsloth/codegemma-7b-it-bnb-4bit" : (
"unsloth/codegemma-7b-it",
"google/codegemma-7b-it",
),
"unsloth/llama-3-8b-bnb-4bit" : (
"unsloth/llama-3-8b",
"meta-llama/Meta-Llama-3-8B",
),
"unsloth/llama-3-8b-Instruct-bnb-4bit" : (
"unsloth/llama-3-8b-Instruct",
"meta-llama/Meta-Llama-3-8B-Instruct",
),
"unsloth/llama-3-70b-bnb-4bit" : (
"meta-llama/Meta-Llama-3-70B",
),
"unsloth/llama-3-70b-Instruct-bnb-4bit" : (
"meta-llama/Meta-Llama-3-70B-Instruct",
),
"unsloth/Phi-3-mini-4k-instruct-bnb-4bit" : (
"unsloth/Phi-3-mini-4k-instruct",
"microsoft/Phi-3-mini-4k-instruct",
),
"unsloth/mistral-7b-v0.3-bnb-4bit" : (
"unsloth/mistral-7b-v0.3",
"mistralai/Mistral-7B-v0.3",
),
"unsloth/mistral-7b-instruct-v0.3-bnb-4bit" : (
"unsloth/mistral-7b-instruct-v0.3",
"mistralai/Mistral-7B-Instruct-v0.3",
),
"unsloth/Phi-3-medium-4k-instruct-bnb-4bit" : (
"unsloth/Phi-3-medium-4k-instruct",
"microsoft/Phi-3-medium-4k-instruct",
),
"unsloth/Qwen2-0.5B-bnb-4bit" : (
"unsloth/Qwen2-0.5B",
"Qwen/Qwen2-0.5B",
),
"unsloth/Qwen2-0.5B-Instruct-bnb-4bit" : (
"unsloth/Qwen2-0.5B-Instruct",
"Qwen/Qwen2-0.5B-Instruct",
),
"unsloth/Qwen2-1.5B-bnb-4bit" : (
"unsloth/Qwen2-1.5B",
"Qwen/Qwen2-1.5B",
),
"unsloth/Qwen2-1.5B-Instruct-bnb-4bit" : (
"unsloth/Qwen2-1.5B-Instruct",
"Qwen/Qwen2-1.5B-Instruct",
),
"unsloth/Qwen2-7B-bnb-4bit" : (
"unsloth/Qwen2-7B",
"Qwen/Qwen2-7B",
),
"unsloth/Qwen2-7B-Instruct-bnb-4bit" : (
"unsloth/Qwen2-7B-Instruct",
"Qwen/Qwen2-7B-Instruct",
),
"unsloth/Qwen2-70B-bnb-4bit" : (
"Qwen/Qwen2-70B",
),
"unsloth/Qwen2-70B-Instruct-bnb-4bit" : (
"Qwen/Qwen2-70B-Instruct",
),
"mistralai/Codestral-22B-v0.1" : (
"mistral-community/Codestral-22B-v0.1",
),
"unsloth/gemma-2-9b-bnb-4bit" : (
"unsloth/gemma-2-9b",
"google/gemma-2-9b",
),
"unsloth/gemma-2-27b-bnb-4bit" : (
"unsloth/gemma-2-27b",
"google/gemma-2-27b",
),
"unsloth/gemma-2-9b-it-bnb-4bit" : (
"unsloth/gemma-2-9b-it",
"google/gemma-2-9b-it",
),
"unsloth/gemma-2-27b-it-bnb-4bit" : (
"unsloth/gemma-2-27b-it",
"google/gemma-2-27b-it",
),
"unsloth/Phi-3-mini-4k-instruct-v0-bnb-4bit" : ( # Old Phi pre July
"unsloth/Phi-3-mini-4k-instruct-v0",
),
"unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit" : ( # New 12b Mistral models
"unsloth/Mistral-Nemo-Instruct-2407",
"mistralai/Mistral-Nemo-Instruct-2407",
),
"unsloth/Mistral-Nemo-Base-2407-bnb-4bit" : ( # New 12b Mistral models
"unsloth/Mistral-Nemo-Base-2407",
"mistralai/Mistral-Nemo-Base-2407",
),
"unsloth/Meta-Llama-3.1-8B-bnb-4bit" : (
"unsloth/Meta-Llama-3.1-8B",
"meta-llama/Meta-Llama-3.1-8B",
),
"unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit" : (
"unsloth/Meta-Llama-3.1-8B-Instruct",
"meta-llama/Meta-Llama-3.1-8B-Instruct",
),
"unsloth/Meta-Llama-3.1-70B-bnb-4bit" : (
"meta-llama/Meta-Llama-3.1-70B",
),
"unsloth/Meta-Llama-3.1-405B-bnb-4bit" : (
"meta-llama/Meta-Llama-3.1-405B",
),
"unsloth/Meta-Llama-3.1-405B-Instruct-bnb-4bit" : (
"meta-llama/Meta-Llama-3.1-405B-Instruct",
),
"unsloth/Meta-Llama-3.1-70B-Instruct-bnb-4bit" : (
"meta-llama/Meta-Llama-3.1-70B-Instruct",
),
"unsloth/Mistral-Large-Instruct-2407-bnb-4bit" : (
"mistralai/Mistral-Large-Instruct-2407",
),
}
INT_TO_FLOAT_MAPPER = {}
FLOAT_TO_INT_MAPPER = {}
for key, values in __INT_TO_FLOAT_MAPPER.items():
INT_TO_FLOAT_MAPPER[key] = values[0]
for value in values:
FLOAT_TO_INT_MAPPER[value] = key
pass
# Get lowercased
lowered_key = key.lower()
INT_TO_FLOAT_MAPPER[lowered_key] = values[0].lower()
for value in values:
FLOAT_TO_INT_MAPPER[value.lower()] = lowered_key
pass
pass
build/lib/build/lib/build/lib/unsloth/models/mistral.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .llama import *
import os
from ._utils import __version__
from .llama import (
LlamaRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
)
from transformers.models.mistral.modeling_mistral import (
MistralAttention,
MistralDecoderLayer,
MistralModel,
MistralForCausalLM,
)
# For Pytorch 2.1.1
try:
from transformers.models.mistral.modeling_mistral import (
MistralSdpaAttention,
MistralFlashAttention2,
)
except:
MistralSdpaAttention = MistralAttention
MistralFlashAttention2 = MistralAttention
pass
def MistralAttention_fast_forward(
self,
hidden_states: torch.Tensor,
causal_mask: Optional[xformers.attn_bias.BlockDiagonalCausalMask] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: bool = False,
use_cache: bool = False,
padding_mask: Optional[torch.LongTensor] = None,
*args, **kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
# Clear inference
if hasattr(self, "paged_attention"):
del self.paged_attention_K
del self.paged_attention_V
del self.paged_attention
del self.temp_QA
del self.temp_KV
del self.RH_Q
del self.attention
pass
bsz, q_len, _ = hidden_states.size()
n_heads = self.num_heads
n_groups = self.num_key_value_groups
n_kv_heads = self.num_key_value_heads
head_dim = self.head_dim
assert(n_kv_heads * n_groups == n_heads)
Q, K, V = self.apply_qkv(self, hidden_states)
Q = Q.view(bsz, q_len, n_heads, head_dim).transpose(1, 2)
K = K.view(bsz, q_len, n_kv_heads, head_dim).transpose(1, 2)
V = V.view(bsz, q_len, n_kv_heads, head_dim).transpose(1, 2)
kv_seq_len = K.shape[-2]
if past_key_value is not None:
kv_seq_len += past_key_value[0].shape[-2]
# Extend RoPE dynamically to fit in VRAM
self.rotary_emb.extend_rope_embedding(V, seq_len = kv_seq_len)
if position_ids is None:
cos = self.rotary_emb.cos_cached
sin = self.rotary_emb.sin_cached
Q, K = fast_rope_embedding(Q, K, cos, sin)
else:
cos, sin = self.rotary_emb(V, seq_len = kv_seq_len)
Q, K = inplace_rope_embedding(Q, K, cos, sin, position_ids)
pass
if past_key_value is not None:
K = torch.cat([past_key_value[0], K], dim = 2)
V = torch.cat([past_key_value[1], V], dim = 2)
pass
past_key_value = (K, V) if use_cache else None
# Attention module
if (not HAS_FLASH_ATTENTION and attention_mask is None):
# Xformers memory efficient attention
Q = Q.transpose(1, 2)
K = K.transpose(1, 2)
V = V.transpose(1, 2)
K_M = V_M = bsz * kv_seq_len
Q_M = bsz * q_len
has_swa = isinstance(causal_mask, xformers.attn_bias.BlockDiagonalCausalMask)
# Group query attention
K = K .view(bsz, kv_seq_len, n_kv_heads, 1, head_dim)
V = V .view(bsz, kv_seq_len, n_kv_heads, 1, head_dim)
K = K.expand(bsz, kv_seq_len, n_kv_heads, n_groups, head_dim)
V = V.expand(bsz, kv_seq_len, n_kv_heads, n_groups, head_dim)
if hidden_states.requires_grad:
K = K.reshape(bsz, kv_seq_len, n_heads, head_dim)
V = V.reshape(bsz, kv_seq_len, n_heads, head_dim)
if has_swa:
Q = Q.view(1, Q_M, n_heads, head_dim)
K = K.view(1, K_M, n_heads, head_dim)
V = V.view(1, V_M, n_heads, head_dim)
pass
else:
# Xformers does support the forward pass though
Q = Q.view(bsz, q_len, n_kv_heads, n_groups, head_dim)
if has_swa:
Q = Q.view(1, Q_M, n_kv_heads, n_groups, head_dim)
K = K.view(1, K_M, n_kv_heads, n_groups, head_dim)
V = V.view(1, V_M, n_kv_heads, n_groups, head_dim)
pass
pass
A = xformers_attention(Q, K, V, attn_bias = causal_mask)
A = A.view(bsz, q_len, n_heads, head_dim)
elif HAS_FLASH_ATTENTION and attention_mask is None:
Q = Q.transpose(1, 2)
K = K.transpose(1, 2)
V = V.transpose(1, 2)
sw = getattr(self.config, "sliding_window", None)
sw = kv_seq_len if (sw is None or sw == "null") else sw
window = (-1, -1) if (kv_seq_len <= sw) else (sw, sw)
A = flash_attn_func(Q, K, V, causal = True, window_size = window)
else:
# Grouped query attention
# if n_groups != 1:
K = K[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, kv_seq_len, head_dim)
V = V[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, kv_seq_len, head_dim)
K = K.reshape(bsz, n_heads, kv_seq_len, head_dim)
V = V.reshape(bsz, n_heads, kv_seq_len, head_dim)
# pass
# Must be contiguous or else results are False!
# https://github.com/pytorch/pytorch/issues/112577
Q, K, V = Q.contiguous(), K.contiguous(), V.contiguous()
# Needs (batch_size, n_heads, seq_len, head_dim)
# is_casual and attention_mask must not be both set!
A = scaled_dot_product_attention(Q, K, V, attn_mask = attention_mask, is_causal = False)
# Go back to (batch_size, seq_len, n_heads, head_dim)
A = A.transpose(1, 2).contiguous()
pass
attn_output = A.reshape(bsz, q_len, n_heads*head_dim)
attn_output = self.apply_o(self, attn_output)
attn_weights = None
return attn_output, attn_weights, past_key_value
pass
def MistralForCausalLM_fast_forward(
self,
input_ids: torch.LongTensor = None,
causal_mask: Optional[xformers.attn_bias.BlockDiagonalCausalMask] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
*args, **kwargs,
) -> Union[Tuple, CausalLMOutputWithPast]:
if causal_mask is None and past_key_values is None:
bsz, q_len = input_ids.shape
sliding_window = getattr(self.config, "sliding_window", None)
if sliding_window is None or sliding_window == "null" or sliding_window <= 0:
causal_mask = xformers.attn_bias.LowerTriangularMask()
elif q_len <= sliding_window:
causal_mask = xformers.attn_bias.LowerTriangularMask()
else:
# Fix from https://github.com/Rypo
causal_mask = xformers.attn_bias.BlockDiagonalCausalMask\
.from_seqlens([q_len]*bsz)\
.make_local_attention(window_size = sliding_window)
pass
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
self.model._has_no_labels = labels is None
if past_key_values is not None:
outputs = LlamaModel_fast_forward_inference(
self,
input_ids,
past_key_values,
position_ids = position_ids,
attention_mask = attention_mask,
)
else:
outputs = self.model(
input_ids=input_ids,
causal_mask=causal_mask,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pass
hidden_states = outputs[0]
bsz, q_len, hd = hidden_states.shape
lm_head = self.lm_head.weight
if bsz == 1 and q_len == 1:
logits = torch.mv(lm_head, hidden_states.ravel().to(lm_head.dtype))
logits = logits.unsqueeze(0).unsqueeze(0)
else:
logits = self.lm_head(hidden_states.to(lm_head.dtype))
pass
logits = logits.to(self.config.torch_dtype)
loss = None
if labels is not None:
shift_logits = logits
if not hasattr(self, "extra_ignored_labels"):
# Fixes https://github.com/unslothai/unsloth/issues/10
self.extra_ignored_labels = torch.full((self.max_seq_length, 1), -100, device = "cuda:0")
pass
shift_labels = torch.hstack((labels[..., 1:], self.extra_ignored_labels[:labels.shape[0]]))
loss = fast_cross_entropy_loss(
logits = shift_logits,
labels = shift_labels,
)
pass
if not return_dict:
output = (logits,) + outputs[1:]
return (loss,) + output if loss is not None else output
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
pass
# Transformers had to update for Mistral Nemo 12b since Attention is (5120, 4096) now.
def patch_mistral_nemo_attention(function):
function = function.replace(
"(self.head_dim * self.num_heads) != self.hidden_size",
"False",
)
function = function.replace(
"self.head_dim = self.hidden_size // self.num_heads",
"self.head_dim = config.head_dim",
)
function = function.replace(
"self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)",
"self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)",
)
return function
pass
class FastMistralModel(FastLlamaModel):
@staticmethod
def pre_patch():
init_name, function = patch_linear_scaling(
model_name = "mistral",
rope_module = LlamaRotaryEmbedding,
scaled_rope_module = LlamaLinearScalingRotaryEmbedding,
attention_module = MistralAttention,
)
# Just for Mistral Nemo models!
if function is not None:
function = patch_mistral_nemo_attention(function)
# if True:#init_name is not None:
exec(function, globals())
MistralAttention.__init__ = eval(init_name)
pass
MistralAttention .forward = MistralAttention_fast_forward
MistralSdpaAttention .forward = MistralAttention_fast_forward
MistralFlashAttention2.forward = MistralAttention_fast_forward
MistralDecoderLayer .forward = LlamaDecoderLayer_fast_forward
MistralModel .forward = LlamaModel_fast_forward
MistralForCausalLM .forward = MistralForCausalLM_fast_forward
PeftModelForCausalLM .forward = PeftModelForCausalLM_fast_forward
fix_prepare_inputs_for_generation(MistralForCausalLM)
# Solves https://github.com/unslothai/unsloth/issues/168
# Static KV Cache was introduced in 4.38.0, causing training to be much slower.
# Inferene can now be CUDAGraphed, but we shall retain the old rotary embeddings.
# https://github.com/huggingface/transformers/pull/27931
# https://github.com/huggingface/transformers/blob/v4.37.2/src/transformers/models/llama/modeling_llama.py
import transformers.models.mistral.modeling_mistral
transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding = LlamaRotaryEmbedding
return
pass
@staticmethod
def from_pretrained(
model_name = "unsloth/mistral-7b-bnb-4bit",
max_seq_length = None,
dtype = None,
load_in_4bit = True,
token = None,
device_map = "sequential",
rope_scaling = None, # Mistral does not support RoPE scaling
fix_tokenizer = True,
model_patcher = None,
tokenizer_name = None,
trust_remote_code = False,
**kwargs,
):
return FastLlamaModel.from_pretrained(
model_name = model_name,
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
token = token,
device_map = device_map,
rope_scaling = rope_scaling,
fix_tokenizer = fix_tokenizer,
model_patcher = FastMistralModel,
tokenizer_name = tokenizer_name,
trust_remote_code = trust_remote_code,
**kwargs,
)
pass
pass
build/lib/build/lib/build/lib/unsloth/models/qwen2.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .llama import *
from .llama import (
LlamaRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
)
from transformers.models.qwen2.modeling_qwen2 import (
Qwen2Attention,
Qwen2DecoderLayer,
Qwen2Model,
Qwen2ForCausalLM,
)
# For Pytorch 2.1.1
try:
from transformers.models.qwen2.modeling_qwen2 import (
Qwen2SdpaAttention,
Qwen2FlashAttention2,
)
except:
Qwen2SdpaAttention = Qwen2Attention
Qwen2FlashAttention2 = Qwen2Attention
pass
class FastQwen2Model(FastLlamaModel):
@staticmethod
def pre_patch():
init_name, function = patch_linear_scaling(
model_name = "qwen2",
rope_module = LlamaRotaryEmbedding,
scaled_rope_module = LlamaLinearScalingRotaryEmbedding,
attention_module = Qwen2Attention,
)
if init_name is not None:
exec(function, globals())
Qwen2Attention.__init__ = eval(init_name)
pass
Qwen2Attention .forward = LlamaAttention_fast_forward
Qwen2SdpaAttention .forward = LlamaAttention_fast_forward
Qwen2FlashAttention2.forward = LlamaAttention_fast_forward
Qwen2DecoderLayer .forward = LlamaDecoderLayer_fast_forward
Qwen2Model .forward = LlamaModel_fast_forward
Qwen2ForCausalLM .forward = CausalLM_fast_forward(LlamaModel_fast_forward_inference)
PeftModelForCausalLM.forward = PeftModelForCausalLM_fast_forward
fix_prepare_inputs_for_generation(Qwen2ForCausalLM)
# Solves https://github.com/unslothai/unsloth/issues/168
# Static KV Cache was introduced in 4.38.0, causing training to be much slower.
# Inferene can now be CUDAGraphed, but we shall retain the old rotary embeddings.
# https://github.com/huggingface/transformers/pull/27931
# https://github.com/huggingface/transformers/blob/v4.37.2/src/transformers/models/llama/modeling_llama.py
import transformers.models.qwen2.modeling_qwen2
transformers.models.qwen2.modeling_qwen2.Qwen2RotaryEmbedding = LlamaRotaryEmbedding
return
pass
@staticmethod
def from_pretrained(
model_name = "Qwen/Qwen2-7B",
max_seq_length = 4096,
dtype = None,
load_in_4bit = True,
token = None,
device_map = "sequential",
rope_scaling = None, # Qwen2 does not support RoPE scaling
fix_tokenizer = True,
model_patcher = None,
tokenizer_name = None,
trust_remote_code = False,
**kwargs,
):
return FastLlamaModel.from_pretrained(
model_name = model_name,
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
token = token,
device_map = device_map,
rope_scaling = rope_scaling,
fix_tokenizer = fix_tokenizer,
model_patcher = FastQwen2Model,
tokenizer_name = tokenizer_name,
trust_remote_code = trust_remote_code,
**kwargs,
)
pass
pass
build/lib/build/lib/build/lib/unsloth/save.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
from peft.tuners.lora import Linear4bit as Peft_Linear4bit
from peft.tuners.lora import Linear as Peft_Linear
from typing import Optional, Callable, Union, List
import torch
import os
import shutil
import pickle
import gc
from transformers.models.llama.modeling_llama import logger
from .kernels import fast_dequantize, QUANT_STATE, get_lora_parameters_bias
import subprocess
import psutil
import re
from transformers.models.llama.modeling_llama import logger
from .tokenizer_utils import fix_sentencepiece_gguf
__all__ = [
"print_quantization_methods",
"unsloth_save_model",
"save_to_gguf",
"patch_saving_functions",
]
# Check environments
keynames = "\n" + "\n".join(os.environ.keys())
IS_COLAB_ENVIRONMENT = "\nCOLAB_" in keynames
IS_KAGGLE_ENVIRONMENT = "\nKAGGLE_" in keynames
del keynames
# Weights
LLAMA_WEIGHTS = (
"self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj", "self_attn.o_proj",
"mlp.gate_proj", "mlp.up_proj", "mlp.down_proj",
)
LLAMA_LAYERNORMS = (
"input_layernorm", "post_attention_layernorm",
"pre_feedforward_layernorm", "post_feedforward_layernorm",
)
# https://github.com/ggerganov/llama.cpp/blob/master/examples/quantize/quantize.cpp#L19
# From https://mlabonne.github.io/blog/posts/Quantize_Llama_2_models_using_ggml.html
ALLOWED_QUANTS = \
{
"not_quantized" : "Recommended. Fast conversion. Slow inference, big files.",
"fast_quantized" : "Recommended. Fast conversion. OK inference, OK file size.",
"quantized" : "Recommended. Slow conversion. Fast inference, small files.",
"f32" : "Not recommended. Retains 100% accuracy, but super slow and memory hungry.",
"bf16" : "Bfloat16 - Fastest conversion + retains 100% accuracy. Slow and memory hungry.",
"f16" : "Float16 - Fastest conversion + retains 100% accuracy. Slow and memory hungry.",
"q8_0" : "Fast conversion. High resource use, but generally acceptable.",
"q4_k_m" : "Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K",
"q5_k_m" : "Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K",
"q2_k" : "Uses Q4_K for the attention.vw and feed_forward.w2 tensors, Q2_K for the other tensors.",
"q3_k_l" : "Uses Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K",
"q3_k_m" : "Uses Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K",
"q3_k_s" : "Uses Q3_K for all tensors",
"q4_0" : "Original quant method, 4-bit.",
"q4_1" : "Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models.",
"q4_k_s" : "Uses Q4_K for all tensors",
"q4_k" : "alias for q4_k_m",
"q5_k" : "alias for q5_k_m",
"q5_0" : "Higher accuracy, higher resource usage and slower inference.",
"q5_1" : "Even higher accuracy, resource usage and slower inference.",
"q5_k_s" : "Uses Q5_K for all tensors",
"q6_k" : "Uses Q8_K for all tensors",
# "iq2_xxs" : "2.06 bpw quantization", # Not supported sadly
# "iq2_xs" : "2.31 bpw quantization",
# "iq3_xxs" : "3.06 bpw quantization",
"q3_k_xs" : "3-bit extra small quantization",
}
def print_quantization_methods():
for key, value in ALLOWED_QUANTS.items():
print(f'"{key}" ==> {value}')
pass
pass
def check_if_sentencepiece_model(model, temporary_location = "_unsloth_sentencepiece_temp"):
if not hasattr(model, "_saved_temp_tokenizer"): return False
temp_tokenizer = model._saved_temp_tokenizer
sentencepiece_model = False
file_location = os.path.join(temporary_location, temp_tokenizer.name_or_path)
if not os.path.exists(file_location):
os.makedirs(file_location)
pass
temp_tokenizer.save_pretrained(file_location)
if os.path.isfile(f"{file_location}/tokenizer.model"):
sentencepiece_model = True
pass
shutil.rmtree(file_location, ignore_errors = True)
return sentencepiece_model
pass
def _free_cached_model(model):
from huggingface_hub import scan_cache_dir
cached_repos = list(scan_cache_dir().repos)
# Go through every cached repo, and delete the one that matches the model we want to save.
# Can save 4GB of disk space - useful for Kaggle systems.
for cached_repo in cached_repos:
if cached_repo.repo_id == model.config._name_or_path:
remove_cache_commit = list(cached_repo.revisions)[0].commit_hash
delete_strategy = scan_cache_dir().delete_revisions(remove_cache_commit,)
logger.warning_once(
"Unsloth: Will remove a cached repo with size " + \
delete_strategy.expected_freed_size_str,
)
delete_strategy.execute()
pass
pass
pass
def _merge_lora(layer, name):
bias = None
if isinstance(layer, (Bnb_Linear4bit, Peft_Linear4bit, Peft_Linear)):
# Is LoRA so we need to merge!
W, quant_state, A, B, s, bias = get_lora_parameters_bias(layer)
if quant_state is not None:
dtype = quant_state.dtype if type(quant_state) is not list else quant_state[2]
W = fast_dequantize(W, quant_state)
else:
dtype = W.dtype
W = W.to(torch.float32).t()
# W = W.t()
if A is not None:
# sAB = (A.t().to(torch.float32) @ (s * B.t().to(torch.float32)))
# W += sAB
W.addmm_(A.t().to(torch.float32), B.t().to(torch.float32), alpha = s)
# W.addmm_(A.t().to(W.dtype), B.t().to(W.dtype), alpha = s)
# if not torch.isfinite(W).all():
maximum_element = torch.max(W.min().abs(), W.max())
if not torch.isfinite(maximum_element).item():
raise ValueError(f"Unsloth: Merge failed.\n{name} has some elements = infinity.")
pass
W = W.t().to(dtype)
else:
W = layer.weight
return W, bias
pass
def fast_save_pickle(shard, name):
# Use this if # CPUs is <= 2
print(f"Unsloth: Saving {name}...")
torch.save(
shard,
name,
# HIGHEST_PROTOCOL seems to not work with Pytorch!
# pickle_module = pickle,
# pickle_protocol = pickle.HIGHEST_PROTOCOL,
)
return
pass
@torch.inference_mode
def unsloth_save_model(
model,
tokenizer,
save_directory : Union[str, os.PathLike],
save_method : str = "lora", # ["lora", "merged_16bit", "merged_4bit"]
push_to_hub : bool = False,
token : Optional[Union[str, bool]] = None,
is_main_process : bool = True,
state_dict : Optional[dict] = None,
save_function : Callable = torch.save,
max_shard_size : Union[int, str] = "5GB",
safe_serialization : bool = True,
variant : Optional[str] = None,
save_peft_format : bool = True,
# Push to hub
use_temp_dir : Optional[bool] = None,
commit_message : Optional[str] = "Trained with Unsloth",
private : Optional[bool] = None,
create_pr : bool = False,
revision : str = None,
commit_description : str = "Upload model trained with Unsloth 2x faster",
tags : List[str] = None,
# Our functions
temporary_location : str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage : float = 0.9,
):
if token is None and "HF_TOKEN" in os.environ:
token = os.environ["HF_TOKEN"]
if token is None and "HUGGINGFACE_TOKEN" in os.environ:
token = os.environ["HUGGINGFACE_TOKEN"]
if commit_message is None: commit_message = ""
if "Unsloth" not in commit_message:
commit_message += " (Trained with Unsloth)"
commit_message = commit_message.lstrip()
if commit_description is None:
commit_description = "Upload model trained with Unsloth 2x faster"
elif "Unsloth 2x faster" not in commit_description:
commit_description += " (Trained with Unsloth 2x faster)"
pass
if save_method == "merged_4bit":
raise RuntimeError(
"Unsloth: Merging into 4bit will cause your model to lose accuracy if you plan\n"\
"to merge to GGUF or others later on. I suggest you to do this as a final step\n"\
"if you're planning to do multiple saves.\n"\
"If you are certain, change `save_method` to `merged_4bit_forced`."
)
elif save_method == "merged_4bit_forced":
save_method = "merged_4bit"
pass
save_pretrained_settings = dict(locals())
for deletion in ("model", "tokenizer", "save_method", "temporary_location", "maximum_memory_usage"):
del save_pretrained_settings[deletion]
pass
# First check for a token!
if push_to_hub:
from huggingface_hub import whoami
try:
username = whoami(token = token)["name"]
except:
raise RuntimeError(
"Unsloth: Please supply a token!\n"\
"Go to https://huggingface.co/settings/tokens"
)
pass
pass
assert(maximum_memory_usage > 0 and maximum_memory_usage <= 0.95)
# Clean memory up first
for _ in range(3):
torch.cuda.empty_cache()
gc.collect()
pass
save_method = save_method.lower().replace(" ", "_")
if save_method != "lora" and save_method != "merged_16bit" and save_method != "merged_4bit":
raise RuntimeError(
"Unsloth: You must select one of 3 options when saving models:\n"\
'"lora" ==> This is the fastest and easiet. Just saves LoRA modules.\n'\
'"merged_16bit" ==> This merges LoRA weights and saves to float16. Needed for llama.cpp / GGUF.\n'\
'"merged_4bit" ==> This merges LoRA weights and saves to 4bit. Useful for DPO / inference.'
)
pass
if save_method == "merged_4bit":
print("Unsloth: Merging 4bit and LoRA weights to 4bit...")
print("This might take 5 minutes...")
# Counteract no LoRA adapters!
if hasattr(model, "merge_and_unload"):
model = model.merge_and_unload()
pass
print("Done.")
pass
if tags is not None:
assert(isinstance(tags, (list, tuple)))
tags = list(tags) + ["unsloth",]
else:
tags = ["unsloth",]
pass
save_pretrained_settings["tags"] = tags
if ((save_method == "lora") or (save_method == "merged_4bit")) and push_to_hub:
if token is None:
raise RuntimeError(
"Unsloth: Pushing to HF requires a token. Pass `token = 'hf_....'`\n"\
"Go to https://huggingface.co/settings/tokens."
)
pass
if save_method == "lora":
print("Unsloth: Saving LoRA adapters. Please wait...")
elif save_method == "merged_4bit":
print("Unsloth: Saving 4bit Bitsandbytes model. Please wait...")
pass
# Update model tag
_ = upload_to_huggingface(
model, save_directory, token,
"finetuned", "trl", file_location = None,
old_username = None, private = private,
)
getattr(model, "original_push_to_hub", tokenizer.push_to_hub)\
(
repo_id = save_directory,
use_temp_dir = use_temp_dir,
commit_message = commit_message,
private = private,
token = token,
max_shard_size = max_shard_size,
create_pr = create_pr,
safe_serialization = safe_serialization,
revision = revision,
commit_description = commit_description,
tags = tags,
)
if tokenizer is not None:
# Set padding side to left for inference
old_padding_side = tokenizer.padding_side
tokenizer.padding_side = "left"
getattr(tokenizer, "original_push_to_hub", tokenizer.push_to_hub)\
(
repo_id = save_directory,
use_temp_dir = use_temp_dir,
commit_message = commit_message,
private = private,
token = token,
max_shard_size = max_shard_size,
create_pr = create_pr,
safe_serialization = safe_serialization,
revision = revision,
commit_description = commit_description,
tags = tags,
)
# Revert back padding side
tokenizer.padding_side = old_padding_side
pass
if hasattr(model, "config"):
print(f"Saved {save_method} model to https://huggingface.co/" + save_directory)
pass
return save_directory, None
pass
# Tokenizer has different saving arguments
tokenizer_save_settings = \
{
"save_directory" : save_pretrained_settings["save_directory"],
"legacy_format" : None,
"filename_prefix" : None,
"push_to_hub" : save_pretrained_settings["push_to_hub"],
"private" : save_pretrained_settings["private"],
"token" : save_pretrained_settings["token"],
}
# Check if PEFT Model or not - if yes, 3 levels. If not 2 levels.
from peft import PeftModelForCausalLM
if isinstance(model, PeftModelForCausalLM):
internal_model = model.model
else:
internal_model = model
pass
# Cannot be converted properly!
if (save_method == "merged_4bit") or (save_method == "lora") or (
not hasattr(model, "model") or \
not hasattr(internal_model.model, "layers")
):
# Do general saving
# Edit save_pretrained_settings
# [TODO] _create_repo has errors due to **kwargs getting accepted
# commit_description does not seem to work?
what_to_delete = ("use_temp_dir", "commit_message", "create_pr", "revision", "commit_description", "tags",) \
if save_pretrained_settings["push_to_hub"] is False else \
("use_temp_dir", "create_pr", "revision", "tags", "commit_description",)
for deletion in what_to_delete:
del save_pretrained_settings[deletion]
pass
if hasattr(model, "add_model_tags"):
model.add_model_tags(["unsloth",])
# Update model tag
if push_to_hub:
_ = upload_to_huggingface(
model, save_pretrained_settings["save_directory"], token,
"finetuned", "trl", file_location = None,
old_username = None, private = private,
)
pass
if tokenizer is not None:
print("Unsloth: Saving tokenizer...", end = "")
# Set padding side to left for inference
old_padding_side = tokenizer.padding_side
tokenizer.padding_side = "left"
tokenizer.save_pretrained(**tokenizer_save_settings)
# Revert back padding side
tokenizer.padding_side = old_padding_side
print(" Done.")
else:
print()
print("Unsloth: Saving model...", end = "")
if save_method != "lora": print(" This might take 10 minutes for Llama-7b...", end = "")
# [TODO] Is this correct?
if save_method == "lora":
save_pretrained_settings["selected_adapters"] = None
pass
model.save_pretrained(**save_pretrained_settings)
if push_to_hub and hasattr(model, "config"):
print("Saved to https://huggingface.co/" + save_pretrained_settings["save_directory"])
pass
print(" Done.")
return save_directory, None
pass
# If push_to_hub, we must remove the .../ part of a repo
username = None
if push_to_hub and "/" in save_directory:
# +1 solves absolute path issues
username = save_directory[:save_directory.find("/")]
new_save_directory = save_directory[save_directory.find("/")+1:]
logger.warning_once(
f"Unsloth: You are pushing to hub, but you passed your HF username = {username}.\n"\
f"We shall truncate {save_directory} to {new_save_directory}"
)
save_pretrained_settings["save_directory"] = new_save_directory
tokenizer_save_settings ["save_directory"] = new_save_directory
save_directory = new_save_directory
pass
print("Unsloth: Merging 4bit and LoRA weights to 16bit...")
# Determine max RAM usage minus sharding
max_ram = psutil.virtual_memory().available
sharded_ram_usage = 5 * 1024 * 1024 * 1024
if type(max_shard_size) is str:
gb_found = re.match("([0-9]{1,})[\s]{0,}GB", max_shard_size, flags = re.IGNORECASE)
mb_found = re.match("([0-9]{1,})[\s]{0,}MB", max_shard_size, flags = re.IGNORECASE)
if gb_found: sharded_ram_usage = int(gb_found.group(1)) * 1024 * 1024 * 1024
elif mb_found: sharded_ram_usage = int(mb_found.group(1)) * 1024 * 1024
elif type(max_shard_size) is int:
sharded_ram_usage = sharded_ram_usage
pass
# Switch to our fast saving modules if it's a slow PC!
n_cpus = psutil.cpu_count(logical = False)
if n_cpus is None: n_cpus = psutil.cpu_count()
if n_cpus is None: n_cpus = 1
if safe_serialization is None:
safe_serialization = True
save_pretrained_settings["safe_serialization"] = safe_serialization
elif safe_serialization and (n_cpus <= 2):
logger.warning_once(
f"Unsloth: You have {n_cpus} CPUs. Using `safe_serialization` is 10x slower.\n"\
f"We shall switch to Pytorch saving, which will take 3 minutes and not 30 minutes.\n"\
f"To force `safe_serialization`, set it to `None` instead.",
)
safe_serialization = False
save_function = fast_save_pickle
save_pretrained_settings["safe_serialization"] = safe_serialization
save_pretrained_settings["save_function"] = save_function
pass
# Only safe_serialization uses more RAM
if safe_serialization:
max_ram -= sharded_ram_usage
else:
max_ram -= sharded_ram_usage*0.25 # Uses much less
pass
max_ram = int(max(0, max_ram) * maximum_memory_usage)
print(f"Unsloth: Will use up to "\
f"{round(max_ram/1024/1024/1024, 2)} out of "\
f"{round(psutil.virtual_memory().total/1024/1024/1024, 2)} RAM for saving.")
# Max directory for disk saving
if not os.path.exists(temporary_location):
os.makedirs(temporary_location)
pass
# Check if Kaggle or Colab, since only 20GB of Disk space allowed.
if IS_KAGGLE_ENVIRONMENT or IS_COLAB_ENVIRONMENT:
# We free up 4GB of space
logger.warning_once(
"Unsloth: Kaggle/Colab has limited disk space. We need to delete the downloaded\n"\
"model which will save 4-16GB of disk space, allowing you to save on Kaggle/Colab."
)
_free_cached_model(internal_model)
pass
# HF also uses a OrderedDict
from collections import OrderedDict
state_dict = OrderedDict()
torch_dtype = internal_model.config.torch_dtype
if type(torch_dtype) is str:
if torch_dtype == "float16": torch_dtype = torch.float16
elif torch_dtype == "bfloat16": torch_dtype = torch.bfloat16
pass
# Check modules to save float32 dtype
state_dict["model.embed_tokens.weight"] = internal_model.model.embed_tokens.weight.data.to(torch_dtype)
max_vram = int(torch.cuda.get_device_properties(0).total_memory * maximum_memory_usage)
from tqdm import tqdm as ProgressBar
for j, layer in enumerate(ProgressBar(internal_model.model.layers)):
for item in LLAMA_WEIGHTS:
proj = eval(f"layer.{item}")
name = f"model.layers.{j}.{item}.weight"
W, bias = _merge_lora(proj, name)
# Bias term
if bias is not None:
state_dict[f"model.layers.{j}.{item}.bias"] = bias
pass
if (torch.cuda.memory_allocated() + W.nbytes) < max_vram:
# Save to GPU memory
state_dict[name] = W
# [TODO] Saving to RAM seems to leak memory???
# elif (max_ram - W.nbytes) > 0:
# # Save to CPU memory
# logger.warning_once(f"We will save to RAM and not VRAM now.")
# state_dict[name] = W.to("cpu", non_blocking = True, copy = True)
# max_ram = max(max_ram - W.nbytes, 0)
else:
# Save to Disk
logger.warning_once(f"We will save to Disk and not RAM now.")
filename = os.path.join(temporary_location, f"{name}.pt")
torch.save(W, filename, pickle_module = pickle, pickle_protocol = pickle.HIGHEST_PROTOCOL,)
state_dict[name] = torch.load(filename, map_location = "cpu", mmap = True)
pass
for item in LLAMA_LAYERNORMS:
try:
# Skip for Gemma 2
state_dict[f"model.layers.{j}.{item}.weight"] = eval(f"layer.{item}.weight.data")
except:
continue
pass
pass
state_dict["model.norm.weight"] = internal_model.model.norm.weight.data
# Check for modules_to_save float32 dtype
# Check for tied weights
if internal_model.model.embed_tokens.weight.data_ptr() != internal_model.lm_head.weight.data_ptr():
state_dict["lm_head.weight"] = internal_model.lm_head.weight.data.to(torch_dtype)
pass
# All tensors MUST be type torch.Tensor and not torch.nn.parameter.Parameter
for key, value in state_dict.items():
if hasattr(value, "data"): state_dict[key] = value = value.data
if type(value) is not torch.Tensor:
logger.warning_once(f"Unsloth: {key} is not a Tensor but a {type(value)}.")
pass
pass
# Edit save_pretrained_settings
# [TODO] _create_repo has errors due to **kwargs getting accepted
save_pretrained_settings["state_dict"] = state_dict
# commit_description does not seem to work?
what_to_delete = ("use_temp_dir", "commit_message", "create_pr", "revision", "commit_description", "tags",) \
if not push_to_hub else \
("use_temp_dir", "create_pr", "revision", "tags", "commit_description",)
for deletion in what_to_delete:
del save_pretrained_settings[deletion]
pass
if hasattr(model, "add_model_tags"):
model.add_model_tags(["unsloth",])
# Update model tag
if push_to_hub:
_ = upload_to_huggingface(
model, save_pretrained_settings["save_directory"], token,
"finetuned", "trl", file_location = None,
old_username = username, private = private,
)
pass
# First check if we're pushing to an organization!
save_directory = save_pretrained_settings["save_directory"]
if save_pretrained_settings["push_to_hub"]:
new_save_directory, new_username = _determine_username(save_directory, username, token)
if token is not None:
from huggingface_hub import whoami
actual_username = whoami(token = token)["name"]
else:
actual_username = username
pass
# Check if pushing to an organization
if save_pretrained_settings["push_to_hub"] and (username != actual_username):
print(f"Unsloth: Saving to organization with address {new_save_directory}")
# We upload everything at the end!
tokenizer_save_settings["push_to_hub"] = False
tokenizer_save_settings["save_directory"] = new_save_directory
pass
# Save tokenizer
if tokenizer is not None:
print("Unsloth: Saving tokenizer...", end = "")
# Set padding side to left for inference
old_padding_side = tokenizer.padding_side
tokenizer.padding_side = "left"
tokenizer.save_pretrained(**tokenizer_save_settings)
# Revert back padding side
tokenizer.padding_side = old_padding_side
print(" Done.")
else:
print()
pass
print("Unsloth: Saving model... This might take 5 minutes for Llama-7b...")
# Since merged, edit quantization_config
old_config = model.config
new_config = model.config.to_dict()
if "quantization_config" in new_config:
del new_config["quantization_config"]
original_model = model
new_config = type(model.config).from_dict(new_config)
while hasattr(original_model, "model"):
original_model = original_model.model
original_model.config = new_config
model.config = new_config
# Save!
# [TODO] --> is this correct?
# save_pretrained_settings["selected_adapters"] = None
# Check if pushing to an organization
if save_pretrained_settings["push_to_hub"] and (username != actual_username):
print(f"Unsloth: Saving to organization with address {new_save_directory}")
# Pushing to organization!
# Sadly .save_pretrained doesn't work :(
# We first save it via .save_pretrained, then upload manually!
save_pretrained_settings["save_directory"] = new_save_directory
save_pretrained_settings["push_to_hub"] = False
internal_model.save_pretrained(**save_pretrained_settings)
# Now manually go through each file and upload them manually!
filenames = os.listdir(new_save_directory)
from huggingface_hub import HfApi
hf_api = HfApi(token = save_pretrained_settings["token"])
print("Unsloth: Uploading all files... Please wait...")
hf_api.upload_folder(
folder_path = new_save_directory,
path_in_repo = ".",
repo_id = new_save_directory,
repo_type = "model",
commit_message = "(Trained with Unsloth)",
ignore_patterns = "*.md",
)
else:
internal_model.save_pretrained(**save_pretrained_settings)
pass
# Revert config back
original_model = model
while hasattr(original_model, "model"):
original_model = original_model.model
original_model.config = old_config
model.config = old_config
print("Done.")
if push_to_hub and hasattr(model, "config"):
print(f"Saved merged model to https://huggingface.co/{username}/{save_directory.lstrip('/')}")
pass
save_pretrained_settings["state_dict"] = None
for j, (key, value) in enumerate(state_dict.items()):
state_dict[key] = None
if j % 10 == 0:
torch.cuda.empty_cache()
gc.collect()
pass
pass
state_dict = None
del state_dict
torch.cuda.empty_cache()
gc.collect()
# Remove temporary location
import shutil
shutil.rmtree(temporary_location, ignore_errors = True)
for _ in range(3):
torch.cuda.empty_cache()
gc.collect()
return save_directory, username
pass
def install_llama_cpp_clone_non_blocking():
full_command = ["git", "clone", "--recursive", "https://github.com/ggerganov/llama.cpp"]
run_installer = subprocess.Popen(full_command, stdout = subprocess.DEVNULL, stderr = subprocess.STDOUT)
return run_installer
pass
def install_llama_cpp_make_non_blocking():
# https://github.com/ggerganov/llama.cpp/issues/7062
# Weirdly GPU conversion for GGUF breaks??
# env = { **os.environ, "LLAMA_CUDA": "1", }
n_jobs = max(int(psutil.cpu_count()*1.5), 1)
# Force make clean
os.system("make clean -C llama.cpp")
full_command = ["make", "all", "-j"+str(n_jobs), "-C", "llama.cpp"]
# https://github.com/ggerganov/llama.cpp/issues/7062
# Weirdly GPU conversion for GGUF breaks??
# run_installer = subprocess.Popen(full_command, env = env, stdout = subprocess.DEVNULL, stderr = subprocess.STDOUT)
run_installer = subprocess.Popen(full_command, stdout = subprocess.DEVNULL, stderr = subprocess.STDOUT)
return run_installer
pass
def install_python_non_blocking(packages = []):
full_command = ["pip", "install"] + packages
run_installer = subprocess.Popen(full_command, stdout = subprocess.DEVNULL, stderr = subprocess.STDOUT)
return run_installer
pass
def install_llama_cpp_old(version = -10):
# Download the 10th latest release since the latest might be broken!
# FALLBACK mechanism
releases = subprocess.check_output(["git", "ls-remote", "--tags", "https://github.com/ggerganov/llama.cpp.git"])
releases = releases.decode("utf-8").replace("\t", " ").split("\n")
for i, x in enumerate(releases):
if "refs/tags/b" not in x: break
releases = releases[:i]
latest = releases[-1]
version = releases[version].split(" ")[0]
# Check if the llama.cpp exists
if os.path.exists("llama.cpp"):
print(
"**[WARNING]** You have a llama.cpp old directory which is broken.\n"\
"Unsloth will DELETE the broken directory and install a new one.\n"\
"Press CTRL + C / cancel this if this is wrong. We shall wait 10 seconds.\n"
)
import time
for i in range(10):
print(f"**[WARNING]** Deleting llama.cpp directory... {10-i} seconds left.")
time.sleep(1)
import shutil
shutil.rmtree("llama.cpp", ignore_errors = True)
pass
# Clone a specific commit
# Also don't use the GPU!
commands = [
"git clone --recursive https://github.com/ggerganov/llama.cpp",
f"cd llama.cpp && git reset --hard {version} && git clean -df",
"make clean -C llama.cpp",
f"make all -j{psutil.cpu_count()*2} -C llama.cpp",
]
for command in commands:
with subprocess.Popen(command, shell = True, stdout = subprocess.PIPE, stderr = subprocess.STDOUT, bufsize = 1) as sp:
for line in sp.stdout:
line = line.decode("utf-8", errors = "replace")
if "undefined reference" in line:
raise RuntimeError("Failed compiling llama.cpp. Please report this ASAP!")
print(line, flush = True, end = "")
pass
pass
# Check if successful
if not os.path.exists("llama.cpp/quantize") and not os.path.exists("llama.cpp/llama-quantize"):
raise RuntimeError(
"Unsloth: The file 'llama.cpp/llama-quantize' or `llama.cpp/quantize` does not exist.\n"\
"But we expect this file to exist! Maybe the llama.cpp developers changed the name?"
)
pass
pass
def install_llama_cpp_blocking(use_cuda = False):
# https://github.com/ggerganov/llama.cpp/issues/7062
# Weirdly GPU conversion for GGUF breaks??
# use_cuda = "LLAMA_CUDA=1" if use_cuda else ""
commands = [
"git clone --recursive https://github.com/ggerganov/llama.cpp",
"make clean -C llama.cpp",
# https://github.com/ggerganov/llama.cpp/issues/7062
# Weirdly GPU conversion for GGUF breaks??
# f"{use_cuda} make all -j{psutil.cpu_count()*2} -C llama.cpp",
f"make all -j{psutil.cpu_count()*2} -C llama.cpp",
"pip install gguf protobuf",
]
if os.path.exists("llama.cpp"): return
for command in commands:
with subprocess.Popen(command, shell = True, stdout = subprocess.PIPE, stderr = subprocess.STDOUT, bufsize = 1) as sp:
for line in sp.stdout:
line = line.decode("utf-8", errors = "replace")
if "undefined reference" in line:
raise RuntimeError("Failed compiling llama.cpp. Please report this ASAP!")
print(line, flush = True, end = "")
pass
pass
pass
def get_executable(executables):
# Get system locations (System Path).split(system separator)
system_directories = os.environ.get("PATH").split(os.pathsep)
for directory in system_directories:
for executable in executables:
path = os.path.join(directory, executable)
# Check if the executable exists and is executable
if os.path.exists(path) and os.access(path, os.X_OK): return path
pass
pass
return None
pass
def save_to_gguf(
model_type : str,
model_dtype : str,
is_sentencepiece : bool = False,
model_directory : str = "unsloth_finetuned_model",
quantization_method = "fast_quantized", # Can be a list of options! ["q4_k_m", "q8_0", "q5_k_m"]
first_conversion : str = None,
_run_installer = None, # Non blocking install of llama.cpp
):
# logger.warning(
# "NOTICE: llama.cpp GGUF conversion is currently unstable, since llama.cpp is\n"\
# "undergoing some major bug fixes as at 5th of May 2024. This is not an Unsloth issue.\n"\
# "Please be patient - GGUF saving should still work, but might not work as well."
# )
assert(model_dtype == "float16" or model_dtype == "bfloat16")
model_dtype = "f16" if model_dtype == "float16" else "bf16"
# Convert quantization_method to list
if isinstance(quantization_method, list): pass
elif isinstance(quantization_method, str): quantization_method = [ quantization_method, ]
elif isinstance(quantization_method, tuple): quantization_method = list(quantization_method)
else:
raise TypeError("Unsloth: quantization_method can only be a string or a list of strings")
pass
# Check if bfloat16 is supported
if model_dtype == "bf16" and not torch.cuda.is_bf16_supported():
logger.warning(
"Unsloth: Cannot convert to bf16 GGUF since your computer doesn't support it.\n"\
"We shall switch instead to f16."
)
model_dtype = "f16"
pass
# Check first_conversion as well
if first_conversion is None:
first_conversion = model_dtype
pass
# Check I quants
for quant_method in quantization_method:
if quant_method.startswith("iq2"):
raise RuntimeError("Unsloth: Currently iq2 type quantizations aren't supported yet - sorry!")
pass
# Careful convert.py is only for Llama / Mistral based archs
use_fast_convert = False
if not is_sentencepiece: use_fast_convert = False # Llama-3
elif model_type == "llama": use_fast_convert = True
elif model_type == "mistral": use_fast_convert = True
pass
logger.warning_once(f"Unsloth: Converting {model_type} model. Can use fast conversion = {use_fast_convert}.")
# Map quant methods
new_quantization_method = []
for quant_method in quantization_method:
if quant_method == "not_quantized": quant_method = model_dtype
elif quant_method == "fast_quantized": quant_method = "q8_0"
elif quant_method == "quantized": quant_method = "q4_k_m"
elif quant_method is None: quant_method = "q8_0"
# Check if wrong method
if quant_method not in ALLOWED_QUANTS.keys():
error = f"Unsloth: Quant method = [{quant_method}] not supported. Choose from below:\n"
for key, value in ALLOWED_QUANTS.items():
error += f"[{key}] => {value}\n"
raise RuntimeError(error)
pass
new_quantization_method.append(quant_method)
pass
quantization_method = new_quantization_method
print_info = \
f"==((====))== Unsloth: Conversion from QLoRA to GGUF information\n"\
f" \\\ /| [0] Installing llama.cpp will take 3 minutes.\n"\
f"O^O/ \_/ \\ [1] Converting HF to GGUF 16bits will take 3 minutes.\n"\
f"\ / [2] Converting GGUF 16bits to {quantization_method} will take 10 minutes each.\n"\
f' "-____-" In total, you will have to wait at least 16 minutes.\n'
print(print_info)
# Check first_conversion format
if first_conversion == "f16" : pass
elif first_conversion == "bf16" : pass
elif first_conversion == "f32" : pass
elif first_conversion == "q8_0" : pass
else:
raise RuntimeError(
f"Unsloth: `first_conversion` can only be one of ['f16', 'bf16', 'f32', 'q8_0'] and not `{first_conversion}`."
)
pass
# Determine whether the system already has llama.cpp installed and the scripts are executable
quantize_location = get_executable(["llama-quantize", "quantize"])
convert_location = get_executable(["convert-hf-to-gguf.py", "convert_hf_to_gguf.py"])
if quantize_location is not None and convert_location is not None:
print("Unsloth: llama.cpp found in the system. We shall skip installation.")
else:
print("Unsloth: [0] Installing llama.cpp. This will take 3 minutes...")
if _run_installer is not None:
error = _run_installer.wait()
else:
error = 0
install_llama_cpp_blocking()
pass
# Check if successful. If not install 10th latest release
# Careful llama.cpp/quantize changed to llama.cpp/llama-quantize
# and llama.cpp/main changed to llama.cpp/llama-cli
# See https://github.com/ggerganov/llama.cpp/pull/7809
quantize_location = None
if os.path.exists("llama.cpp/quantize"):
quantize_location = "llama.cpp/quantize"
elif os.path.exists("llama.cpp/llama-quantize"):
quantize_location = "llama.cpp/llama-quantize"
else:
raise RuntimeError(
"Unsloth: The file 'llama.cpp/llama-quantize' or 'llama.cpp/quantize' does not exist.\n"\
"But we expect this file to exist! Maybe the llama.cpp developers changed the name?"
)
pass
# See https://github.com/unslothai/unsloth/pull/730
# Filenames changed again!
convert_location = None
if os.path.exists("llama.cpp/convert-hf-to-gguf.py"):
convert_location = "llama.cpp/convert-hf-to-gguf.py"
elif os.path.exists("llama.cpp/convert_hf_to_gguf.py"):
convert_location = "llama.cpp/convert_hf_to_gguf.py"
else:
raise RuntimeError(
"Unsloth: The file 'llama.cpp/convert-hf-to-gguf.py' or 'llama.cpp/convert_hf_to_gguf.py' does not exist.\n"\
"But we expect this file to exist! Maybe the llama.cpp developers changed the name?"
)
pass
if error != 0 or quantize_location is None or convert_location is None:
print(f"Unsloth: llama.cpp error code = {error}.")
install_llama_cpp_old(-10)
pass
pass
# Determine maximum first_conversion state
if first_conversion == "f32" : strength = 3
elif first_conversion == "f16" : strength = 2
elif first_conversion == "bf16" : strength = 1
elif first_conversion == "q8_0" : strength = 0
for quant_method in quantization_method:
if quant_method == "f32": strength = max(strength, 3)
elif quant_method == "f16": strength = max(strength, 2)
elif quant_method == "bf16": strength = max(strength, 1)
elif quant_method == "q8_0": strength = max(strength, 0)
else:
# Quantized models must have f16 as the default argument
if first_conversion == "f32" : pass
elif first_conversion == "f16" : pass
elif first_conversion == "bf16" : pass
elif first_conversion == "q8_0":
logger.warning_once(
"Unsloth: Using q8_0 for the `first_conversion` will lose a bit of accuracy, "\
"but saves disk space!"
)
# first_conversion = "f16"
pass
pass
pass
# If only q8_0:
if len(quantization_method) == 1 and quantization_method[0] == "q8_0":
strength = 0
pass
if strength >= 3: first_conversion = "f32"
elif strength >= 2: first_conversion = "f16"
elif strength >= 1: first_conversion = "bf16"
else: first_conversion = "q8_0"
# Non llama/mistral needs can only use f32 or f16
if not use_fast_convert and \
(first_conversion != "f16" or first_conversion != "bf16" or first_conversion != "f32"):
pass
# Latest llama.cpp works for all models for q8_0!
# logger.warning_once("Unsloth: We must use f16 for non Llama and Mistral models.")
# first_conversion = "f16"
pass
# Check if bfloat16 is supported
if first_conversion == "bf16" and not torch.cuda.is_bf16_supported():
logger.warning(
"Unsloth: Cannot convert to bf16 GGUF since your computer doesn't support it.\n"\
"We shall switch instead to f16."
)
first_conversion = "f16"
pass
n_cpus = psutil.cpu_count()
if n_cpus is None: n_cpus = 1
n_cpus *= 2
# Concurrency from https://rentry.org/llama-cpp-conversions#merging-loras-into-a-model
final_location = f"./{model_directory}/unsloth.{first_conversion.upper()}.gguf"
print(f"Unsloth: [1] Converting model at {model_directory} into {first_conversion} GGUF format.\n"\
f"The output location will be {final_location}\n"\
"This will take 3 minutes...")
# We first check if tokenizer.model exists in the model_directory
if os.path.exists(f"{model_directory}/tokenizer.model"):
vocab_type = "spm,hfft,bpe"
# Fix Sentencepiece model as well!
fix_sentencepiece_gguf(model_directory)
else:
vocab_type = "bpe"
pass
# convert.py is deprecated!
use_fast_convert = False
if use_fast_convert:
command = f"python llama.cpp/convert.py {model_directory} "\
f"--outfile {final_location} --vocab-type {vocab_type} "\
f"--outtype {first_conversion} --concurrency {n_cpus} --pad-vocab"
else:
command = f"python {convert_location} {model_directory} "\
f"--outfile {final_location} "\
f"--outtype {first_conversion}"
pass
with subprocess.Popen(command, shell = True, stdout = subprocess.PIPE, stderr = subprocess.STDOUT, bufsize = 1) as sp:
for line in sp.stdout:
line = line.decode("utf-8", errors = "replace")
if "undefined reference" in line:
raise RuntimeError("Failed compiling llama.cpp. Please report this ASAP!")
print(line, flush = True, end = "")
if sp.returncode is not None and sp.returncode != 0:
raise subprocess.CalledProcessError(sp.returncode, sp.args)
pass
# Check if quantization succeeded!
if not os.path.isfile(final_location):
if IS_KAGGLE_ENVIRONMENT:
raise RuntimeError(
f"Unsloth: Quantization failed for {final_location}\n"\
"You are in a Kaggle environment, which might be the reason this is failing.\n"\
"Kaggle only provides 20GB of disk space. Merging to 16bit for 7b models use 16GB of space.\n"\
"This means using `model.{save_pretrained/push_to_hub}_merged` works, but\n"\
"`model.{save_pretrained/push_to_hub}_gguf will use too much disk space.\n"\
"I suggest you to save the 16bit model first, then use manual llama.cpp conversion."
)
else:
raise RuntimeError(
f"Unsloth: Quantization failed for {final_location}\n"\
"You might have to compile llama.cpp yourself, then run this again.\n"\
"You do not need to close this Python program. Run the following commands in a new terminal:\n"\
"You must run this in the same folder as you're saving your model.\n"\
"git clone --recursive https://github.com/ggerganov/llama.cpp\n"\
"cd llama.cpp && make clean && make all -j\n"\
"Once that's done, redo the quantization."
)
pass
pass
print(f"Unsloth: Conversion completed! Output location: {final_location}")
full_precision_location = final_location
all_saved_locations = [full_precision_location,]
# Convert each type!
for quant_method in quantization_method:
if quant_method != first_conversion:
print(f"Unsloth: [2] Converting GGUF 16bit into {quant_method}. This will take 20 minutes...")
final_location = f"./{model_directory}/unsloth.{quant_method.upper()}.gguf"
command = f"./{quantize_location} {full_precision_location} "\
f"{final_location} {quant_method} {n_cpus}"
# quantize uses stderr
with subprocess.Popen(command, shell = True, stdout = subprocess.PIPE, stderr = subprocess.STDOUT, bufsize = 1) as sp:
for line in sp.stdout:
line = line.decode("utf-8", errors = "replace")
if "undefined reference" in line:
raise RuntimeError("Failed compiling llama.cpp. Please report this ASAP!")
print(line, flush = True, end = "")
if sp.returncode is not None and sp.returncode != 0:
raise subprocess.CalledProcessError(sp.returncode, sp.args)
pass
# Check if quantization succeeded!
if not os.path.isfile(final_location):
if IS_KAGGLE_ENVIRONMENT:
raise RuntimeError(
f"Unsloth: Quantization failed for {final_location}\n"\
"You are in a Kaggle environment, which might be the reason this is failing.\n"\
"Kaggle only provides 20GB of disk space. Merging to 16bit for 7b models use 16GB of space.\n"\
"This means using `model.{save_pretrained/push_to_hub}_merged` works, but\n"\
"`model.{save_pretrained/push_to_hub}_gguf will use too much disk space.\n"\
"I suggest you to save the 16bit model first, then use manual llama.cpp conversion."
)
else:
raise RuntimeError(
"Unsloth: Quantization failed! You might have to compile llama.cpp yourself, then run this again.\n"\
"You do not need to close this Python program. Run the following commands in a new terminal:\n"\
"You must run this in the same folder as you're saving your model.\n"\
"git clone --recursive https://github.com/ggerganov/llama.cpp\n"\
"cd llama.cpp && make clean && make all -j\n"\
"Once that's done, redo the quantization."
)
pass
pass
print(f"Unsloth: Conversion completed! Output location: {final_location}")
all_saved_locations.append(final_location)
pass
pass
return all_saved_locations
pass
def unsloth_save_pretrained_merged(
self,
save_directory : Union[str, os.PathLike],
tokenizer = None,
save_method : str = "merged_16bit", # ["lora", "merged_16bit", "merged_4bit"]
push_to_hub : bool = False,
token : Optional[Union[str, bool]] = None,
is_main_process : bool = True,
state_dict : Optional[dict] = None,
save_function : Callable = torch.save,
max_shard_size : Union[int, str] = "5GB",
safe_serialization : bool = True,
variant : Optional[str] = None,
save_peft_format : bool = True,
tags : List[str] = None,
temporary_location : str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage : float = 0.75,
):
"""
Same as .save_pretrained(...) except 4bit weights are auto
converted to float16 with as few overhead as possible.
Choose for `save_method` to be either:
1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.
2. `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.
3. `lora`: Save LoRA adapters with no merging. Useful for HF inference.
"""
if tokenizer is None:
logger.warning_once(
"Unsloth: You're not saving a tokenizer as well?\n"\
"You can do it separately via `tokenizer.save_pretrained(...)`"
)
pass
arguments = dict(locals())
arguments["model"] = self
del arguments["self"]
unsloth_save_model(**arguments)
for _ in range(3):
gc.collect()
pass
def unsloth_push_to_hub_merged(
self,
repo_id : str,
tokenizer = None,
save_method : str = "merged_16bit", # ["lora", "merged_16bit", "merged_4bit"]
use_temp_dir : Optional[bool] = None,
commit_message : Optional[str] = "Trained with Unsloth",
private : Optional[bool] = None,
token : Union[bool, str, None] = None,
max_shard_size : Union[int, str, None] = "5GB",
create_pr : bool = False,
safe_serialization : bool = True,
revision : str = None,
commit_description : str = "Upload model trained with Unsloth 2x faster",
tags : Optional[List[str]] = None,
temporary_location : str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage : float = 0.75,
):
"""
Same as .push_to_hub(...) except 4bit weights are auto
converted to float16 with as few overhead as possible.
Choose for `save_method` to be either:
1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.
2. `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.
3. `lora`: Save LoRA adapters with no merging. Useful for HF inference.
"""
if tokenizer is None:
logger.warning_once(
"Unsloth: You're not saving a tokenizer as well?\n"\
"You can do it separately via `tokenizer.push_to_hub(...)`"
)
pass
arguments = dict(locals())
arguments["model"] = self
arguments["save_directory"] = repo_id
arguments["push_to_hub"] = True
del arguments["self"]
del arguments["repo_id"]
unsloth_save_model(**arguments)
for _ in range(3):
gc.collect()
pass
MODEL_CARD = \
"""---
base_model: {base_model}
tags:
- text-generation-inference
- transformers
- unsloth
- {model_type}
- {extra}
license: apache-2.0
language:
- en
---
# Uploaded {method} model
- **Developed by:** {username}
- **License:** apache-2.0
- **Finetuned from model :** {base_model}
This {model_type} model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
"""
def _determine_username(save_directory, old_username, token):
username = ""
save_directory = save_directory.lstrip("./")
if "/" not in save_directory:
from huggingface_hub import whoami
try:
username = whoami(token = token)["name"]
if type(old_username) is str and username != old_username:
username = old_username
pass
save_directory = f"{username}/{save_directory}"
except:
raise RuntimeError(f"Unsloth: {save_directory} is not a Huggingface directory.")
else:
username = save_directory.split("/")[0]
pass
return save_directory, username
pass
def upload_to_huggingface(
model,
save_directory,
token,
method,
extra = "",
file_location = None,
old_username = None,
private = None,
):
save_directory, username = _determine_username(save_directory, old_username, token)
from huggingface_hub import create_repo
try:
create_repo(
repo_id = save_directory,
token = token,
repo_type = "model",
exist_ok = False,
private = private,
)
# Create model card
from huggingface_hub import ModelCard
content = MODEL_CARD.format(
username = username,
base_model = model.config._name_or_path,
model_type = model.config.model_type,
method = "",
extra = extra,
)
card = ModelCard(content)
card.push_to_hub(save_directory, token = token)
except:
pass
if file_location is not None:
# Now upload file
from huggingface_hub import HfApi
hf_api = HfApi(token = token)
if "/" in file_location:
uploaded_location = file_location[file_location.rfind("/")+1:]
else:
uploaded_location = file_location
pass
# find ftevent file from tensorboard and upload it
import glob
ftevent_files = glob.glob("*out.tfevents*", recursive = True)
if len(ftevent_files) > 0:
print("Unsloth: Uploading tensorboard files... Please wait...", file_location + "*out.tfevents*")
for ftevent_file in ftevent_files:
hf_api.upload_file(
path_or_fileobj = ftevent_file,
path_in_repo = ftevent_file.replace(file_location, ""),
repo_id = save_directory,
repo_type = "model",
commit_message = "(Trained with Unsloth)",
)
hf_api.upload_file(
path_or_fileobj = file_location,
path_in_repo = uploaded_location,
repo_id = save_directory,
repo_type = "model",
commit_message = "(Trained with Unsloth)",
)
# We also upload a config.json file
import json
with open("_temporary_unsloth_config.json", "w") as file:
json.dump({"model_type" : model.config.model_type}, file, indent = 4)
pass
hf_api.upload_file(
path_or_fileobj = "_temporary_unsloth_config.json",
path_in_repo = "config.json",
repo_id = save_directory,
repo_type = "model",
commit_message = "(Trained with Unsloth)",
)
os.remove("_temporary_unsloth_config.json")
pass
return username
pass
def fix_tokenizer_bos_token(tokenizer):
# Check if BOS added already, then warn
fix_bos_token = False
chat_template = getattr(tokenizer, "chat_template", None)
if (tokenizer("A").input_ids[0] == getattr(tokenizer, "bos_token_id", None)):
if chat_template is not None and \
(
tokenizer.bos_token in chat_template or \
"{bos_token}" in chat_template.replace(" ", "") or \
"{bos_token+" in chat_template.replace(" ", "")
):
fix_bos_token = True
logger.warning(
f"Unsloth: ##### The current model auto adds a BOS token.\n"\
"Unsloth: ##### Your chat template has a BOS token. We shall remove it temporarily."
)
# Remove {{bos_token}}
new_chat_template = re.sub(r"\{[\s]{0,}\{[\s]{0,}bos\_token[\s]{0,}\}[\s]{0,}\}", "", chat_template)
# Remove {{bos_token +
new_chat_template = re.sub(r"\{[\s]{0,}\{[\s]{0,}bos\_token[\s]{0,}\+[\s]{0,}", "", new_chat_template)
tokenizer.chat_template = new_chat_template
pass
pass
return fix_bos_token, chat_template
pass
def create_ollama_modelfile(tokenizer, gguf_location):
"""
Creates an Ollama Modelfile.
Use ollama.create(model = "new_ollama_model", modelfile = modelfile)
"""
modelfile = getattr(tokenizer, "_ollama_modelfile", None)
if modelfile is None: return None
modelfile = modelfile\
.replace("{{", "⚫@✅#🦥")\
.replace("}}", "⚡@🦥#⛵")\
.format(
__FILE_LOCATION__ = gguf_location,
)\
.replace("⚫@✅#🦥", "{{")\
.replace("⚡@🦥#⛵", "}}")\
.rstrip()
pass
return modelfile
pass
def unsloth_save_pretrained_gguf(
self,
save_directory : Union[str, os.PathLike],
tokenizer = None,
quantization_method : str = "fast_quantized",
first_conversion : str = None,
push_to_hub : bool = False,
token : Optional[Union[str, bool]] = None,
private : Optional[bool] = None,
is_main_process : bool = True,
state_dict : Optional[dict] = None,
save_function : Callable = torch.save,
max_shard_size : Union[int, str] = "5GB",
safe_serialization : bool = True,
variant : Optional[str] = None,
save_peft_format : bool = True,
tags : List[str] = None,
temporary_location : str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage : float = 0.85,
):
"""
Same as .save_pretrained(...) except 4bit weights are auto
converted to float16 then converted to GGUF / llama.cpp format.
Choose for `quantization_method` to be:
"not_quantized" : "Recommended. Fast conversion. Slow inference, big files.",
"fast_quantized" : "Recommended. Fast conversion. OK inference, OK file size.",
"quantized" : "Recommended. Slow conversion. Fast inference, small files.",
"f32" : "Not recommended. Retains 100% accuracy, but super slow and memory hungry.",
"f16" : "Fastest conversion + retains 100% accuracy. Slow and memory hungry.",
"q8_0" : "Fast conversion. High resource use, but generally acceptable.",
"q4_k_m" : "Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K",
"q5_k_m" : "Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K",
"q2_k" : "Uses Q4_K for the attention.vw and feed_forward.w2 tensors, Q2_K for the other tensors.",
"q3_k_l" : "Uses Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K",
"q3_k_m" : "Uses Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K",
"q3_k_s" : "Uses Q3_K for all tensors",
"q4_0" : "Original quant method, 4-bit.",
"q4_1" : "Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models.",
"q4_k_s" : "Uses Q4_K for all tensors",
"q4_k" : "alias for q4_k_m",
"q5_k" : "alias for q5_k_m",
"q5_0" : "Higher accuracy, higher resource usage and slower inference.",
"q5_1" : "Even higher accuracy, resource usage and slower inference.",
"q5_k_s" : "Uses Q5_K for all tensors",
"q6_k" : "Uses Q8_K for all tensors",
"iq2_xxs" : "2.06 bpw quantization",
"iq2_xs" : "2.31 bpw quantization",
"iq3_xxs" : "3.06 bpw quantization",
"q3_k_xs" : "3-bit extra small quantization",
"""
if tokenizer is None:
raise ValueError("Unsloth: Saving to GGUF must have a tokenizer.")
arguments = dict(locals())
arguments["model"] = self
arguments["tokenizer"] = tokenizer
arguments["push_to_hub"] = False # We save ourselves
arguments["save_method"] = "merged_16bit" # Must be 16bit
del arguments["self"]
del arguments["quantization_method"]
del arguments["first_conversion"]
# Fix tokenizer adding an extra BOS token at the front
fix_bos_token, old_chat_template = fix_tokenizer_bos_token(tokenizer)
# Non blocking install GGUF first
if not os.path.exists("llama.cpp"):
if IS_KAGGLE_ENVIRONMENT:
# Kaggle is weird - no blocking installs, and no CUDA?
python_install = install_python_non_blocking(["gguf", "protobuf"])
python_install.wait()
install_llama_cpp_blocking(use_cuda = False)
new_save_directory, old_username = unsloth_save_model(**arguments)
makefile = None
else:
git_clone = install_llama_cpp_clone_non_blocking()
python_install = install_python_non_blocking(["gguf", "protobuf"])
git_clone.wait()
makefile = install_llama_cpp_make_non_blocking()
new_save_directory, old_username = unsloth_save_model(**arguments)
python_install.wait()
pass
else:
try:
new_save_directory, old_username = unsloth_save_model(**arguments)
makefile = None
except:
# Retry by recloning llama.cpp
if IS_KAGGLE_ENVIRONMENT:
# Kaggle is weird - no blocking installs, and no CUDA?
python_install = install_python_non_blocking(["gguf", "protobuf"])
python_install.wait()
install_llama_cpp_blocking(use_cuda = False)
new_save_directory, old_username = unsloth_save_model(**arguments)
makefile = None
else:
git_clone = install_llama_cpp_clone_non_blocking()
python_install = install_python_non_blocking(["gguf", "protobuf"])
git_clone.wait()
makefile = install_llama_cpp_make_non_blocking()
new_save_directory, old_username = unsloth_save_model(**arguments)
python_install.wait()
pass
pass
pass
# Use old chat template if the bos is removed
if fix_bos_token:
tokenizer.chat_template = old_chat_template
pass
for _ in range(3):
gc.collect()
model_dtype = self.config.torch_dtype
model_type = self.config.model_type
if type(model_dtype) is str:
assert(model_dtype == "float16" or model_dtype == "bfloat16")
elif model_dtype == torch.float16:
model_dtype = "float16"
elif model_dtype == torch.bfloat16:
model_dtype = "bfloat16"
else:
raise TypeError("Unsloth: Model dtype can only be float16 or bfloat16")
pass
is_sentencepiece_model = check_if_sentencepiece_model(self)
# Save to GGUF
all_file_locations = save_to_gguf(model_type, model_dtype, is_sentencepiece_model,
new_save_directory, quantization_method, first_conversion, makefile,
)
# Save Ollama modelfile
modelfile = create_ollama_modelfile(tokenizer, all_file_locations[0])
modelfile_location = None
if modelfile is not None:
modelfile_location = os.path.join(new_save_directory, "Modelfile")
with open(modelfile_location, "w") as file:
file.write(modelfile)
pass
print(f"Unsloth: Saved Ollama Modelfile to {modelfile_location}")
pass
if fix_bos_token:
logger.warning(
f"Unsloth: ##### The current model auto adds a BOS token.\n"\
"Unsloth: ##### We removed it in GGUF's chat template for you."
)
pass
if push_to_hub:
print("Unsloth: Uploading GGUF to Huggingface Hub...")
for file_location in all_file_locations:
username = upload_to_huggingface(
self, save_directory, token,
"GGUF converted", "gguf", file_location, old_username, private,
)
link = f"{username}/{new_save_directory.lstrip('/.')}" \
if username not in new_save_directory else \
new_save_directory.lstrip('/.')
print(f"Saved GGUF to https://huggingface.co/{link}")
pass
# Save modelfile
if modelfile_location is not None:
username = upload_to_huggingface(
self, save_directory, token,
"GGUF converted", "gguf", modelfile_location, old_username, private,
)
print(f"Saved Ollama Modelfile to https://huggingface.co/{link}")
pass
pass
pass
def unsloth_push_to_hub_gguf(
self,
repo_id : str,
tokenizer = None,
quantization_method : str = "fast_quantized",
first_conversion : str = None,
use_temp_dir : Optional[bool] = None,
commit_message : Optional[str] = "Trained with Unsloth",
private : Optional[bool] = None,
token : Union[bool, str, None] = None,
max_shard_size : Union[int, str, None] = "5GB",
create_pr : bool = False,
safe_serialization : bool = True,
revision : str = None,
commit_description : str = "Upload model trained with Unsloth 2x faster",
tags : Optional[List[str]] = None,
temporary_location : str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage : float = 0.85,
):
"""
Same as .push_to_hub(...) except 4bit weights are auto
converted to float16 then converted to GGUF / llama.cpp format.
Choose for `quantization_method` to be:
"not_quantized" : "Recommended. Fast conversion. Slow inference, big files.",
"fast_quantized" : "Recommended. Fast conversion. OK inference, OK file size.",
"quantized" : "Recommended. Slow conversion. Fast inference, small files.",
"f32" : "Not recommended. Retains 100% accuracy, but super slow and memory hungry.",
"f16" : "Fastest conversion + retains 100% accuracy. Slow and memory hungry.",
"q8_0" : "Fast conversion. High resource use, but generally acceptable.",
"q4_k_m" : "Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K",
"q5_k_m" : "Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K",
"q2_k" : "Uses Q4_K for the attention.vw and feed_forward.w2 tensors, Q2_K for the other tensors.",
"q3_k_l" : "Uses Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K",
"q3_k_m" : "Uses Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K",
"q3_k_s" : "Uses Q3_K for all tensors",
"q4_0" : "Original quant method, 4-bit.",
"q4_1" : "Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models.",
"q4_k_s" : "Uses Q4_K for all tensors",
"q5_0" : "Higher accuracy, higher resource usage and slower inference.",
"q5_1" : "Even higher accuracy, resource usage and slower inference.",
"q5_k_s" : "Uses Q5_K for all tensors",
"q6_k" : "Uses Q8_K for all tensors",
"""
if tokenizer is None:
raise ValueError("Unsloth: Saving to GGUF must have a tokenizer.")
arguments = dict(locals())
arguments["model"] = self
arguments["tokenizer"] = tokenizer
arguments["save_directory"] = repo_id
arguments["push_to_hub"] = False # We save ourselves
arguments["save_method"] = "merged_16bit" # Must be 16bit
del arguments["self"]
del arguments["repo_id"]
del arguments["quantization_method"]
del arguments["first_conversion"]
# Fix tokenizer adding an extra BOS token at the front
fix_bos_token, old_chat_template = fix_tokenizer_bos_token(tokenizer)
# Non blocking install GGUF first
if not os.path.exists("llama.cpp"):
if IS_KAGGLE_ENVIRONMENT:
# Kaggle is weird - no blocking installs, and no CUDA?
python_install = install_python_non_blocking(["gguf", "protobuf"])
python_install.wait()
install_llama_cpp_blocking(use_cuda = False)
new_save_directory, old_username = unsloth_save_model(**arguments)
makefile = None
else:
git_clone = install_llama_cpp_clone_non_blocking()
python_install = install_python_non_blocking(["gguf", "protobuf"])
git_clone.wait()
makefile = install_llama_cpp_make_non_blocking()
new_save_directory, old_username = unsloth_save_model(**arguments)
python_install.wait()
pass
else:
try:
new_save_directory, old_username = unsloth_save_model(**arguments)
makefile = None
except:
# Retry by recloning llama.cpp
if IS_KAGGLE_ENVIRONMENT:
# Kaggle is weird - no blocking installs, and no CUDA?
python_install = install_python_non_blocking(["gguf", "protobuf"])
python_install.wait()
install_llama_cpp_blocking(use_cuda = False)
new_save_directory, old_username = unsloth_save_model(**arguments)
makefile = None
else:
git_clone = install_llama_cpp_clone_non_blocking()
python_install = install_python_non_blocking(["gguf", "protobuf"])
git_clone.wait()
makefile = install_llama_cpp_make_non_blocking()
new_save_directory, old_username = unsloth_save_model(**arguments)
python_install.wait()
pass
pass
pass
# Use old chat template if the bos is removed
if fix_bos_token:
tokenizer.chat_template = old_chat_template
pass
for _ in range(3):
gc.collect()
model_dtype = self.config.torch_dtype
model_type = self.config.model_type
if type(model_dtype) is str:
assert(model_dtype == "float16" or model_dtype == "bfloat16")
elif model_dtype == torch.float16:
model_dtype = "float16"
elif model_dtype == torch.bfloat16:
model_dtype = "bfloat16"
else:
raise TypeError("Unsloth: Model dtype can only be float16 or bfloat16")
pass
is_sentencepiece_model = check_if_sentencepiece_model(self)
# Save to GGUF
all_file_locations = save_to_gguf(model_type, model_dtype, is_sentencepiece_model,
new_save_directory, quantization_method, first_conversion, makefile,
)
# Save Ollama modelfile
modelfile = create_ollama_modelfile(tokenizer, all_file_locations[0])
modelfile_location = None
if modelfile is not None:
modelfile_location = os.path.join(new_save_directory, "Modelfile")
with open(modelfile_location, "w") as file:
file.write(modelfile)
pass
print(f"Unsloth: Saved Ollama Modelfile to {modelfile_location}")
pass
for file_location in all_file_locations:
print("Unsloth: Uploading GGUF to Huggingface Hub...")
username = upload_to_huggingface(
self, repo_id, token,
"GGUF converted", "gguf", file_location, old_username, private,
)
link = f"{username}/{new_save_directory.lstrip('/.')}" \
if username not in new_save_directory else \
new_save_directory.lstrip('/.')
print(f"Saved GGUF to https://huggingface.co/{link}")
pass
# Save modelfile
if modelfile_location is not None:
username = upload_to_huggingface(
self, repo_id, token,
"GGUF converted", "gguf", modelfile_location, old_username, private,
)
print(f"Saved Ollama Modelfile to https://huggingface.co/{link}")
pass
if fix_bos_token:
logger.warning(
f"Unsloth: ##### The current model auto adds a BOS token.\n"\
"Unsloth: ##### We removed it in GGUF's chat template for you."
)
pass
pass
# Corrected function to save LoRA to a custom directory
def save_lora_to_custom_dir(model, tokenizer, save_directory):
# Create the custom directory if it doesn't exist
os.makedirs(save_directory, exist_ok=True)
# Call the unsloth_save_model function with the custom directory
unsloth_save_model(
model,
tokenizer,
save_directory=save_directory,
save_method="lora",
push_to_hub=False,
)
# Corrected method within the model class to convert LoRA to GGML and push to Hugging Face Hub
def unsloth_convert_lora_to_ggml_and_push_to_hub(
self,
tokenizer,
repo_id: str,
use_temp_dir: Optional[bool] = None,
commit_message: Optional[str] = "Converted LoRA to GGML with Unsloth",
private: Optional[bool] = None,
token: Union[bool, str, None] = None,
create_pr: bool = False,
revision: str = None,
commit_description: str = "Convert LoRA to GGML format using Unsloth",
temporary_location: str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage: float = 0.85,
):
if not os.path.exists("llama.cpp"):
if IS_KAGGLE_ENVIRONMENT:
python_install = install_python_non_blocking(["protobuf"])
python_install.wait()
install_llama_cpp_blocking(use_cuda=False)
makefile = None
else:
git_clone = install_llama_cpp_clone_non_blocking()
python_install = install_python_non_blocking(["protobuf"])
git_clone.wait()
makefile = install_llama_cpp_make_non_blocking()
python_install.wait()
else:
makefile = None
for _ in range(3):
gc.collect()
lora_directory_push = "lora-to-ggml-push"
save_lora_to_custom_dir(self, tokenizer, lora_directory_push)
model_type = self.config.model_type
output_file = os.path.join(lora_directory_push, "ggml-adapter-model.bin")
print(f"Unsloth: Converting auto-saved LoRA adapters at {lora_directory_push} to GGML format.")
print(f"The output file will be {output_file}")
command = f"python3 llama.cpp/convert-lora-to-ggml.py {lora_directory_push} {output_file} llama"
try:
with subprocess.Popen(command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=1, universal_newlines=True) as sp:
for line in sp.stdout:
print(line, end="", flush=True)
for line in sp.stderr:
print(line, end="", flush=True)
sp.wait()
if sp.returncode != 0:
raise subprocess.CalledProcessError(sp.returncode, command)
except subprocess.CalledProcessError as e:
print(f"Error: Conversion failed with return code {e.returncode}")
return
print(f"Unsloth: Conversion completed! Output file: {output_file}")
print("Unsloth: Uploading GGML file to Hugging Face Hub...")
username = upload_to_huggingface(
self, repo_id, token,
"GGML converted LoRA", "ggml", output_file, None, private,
)
link = f"{repo_id.lstrip('/')}"
print("Unsloth: Done.")
print(f"Converted LoRA to GGML and uploaded to https://huggingface.co/{link}")
print("\nThis GGML making function was made by Maheswar. Ping him @Maheswar on the Unsloth Discord or on HuggingFace (@mahiatlinux) if you like this!")
def unsloth_convert_lora_to_ggml_and_save_locally(
self,
save_directory: str, # Added parameter for the folder name
tokenizer,
temporary_location: str = "_unsloth_temporary_saved_buffers",
maximum_memory_usage: float = 0.85,
):
if not os.path.exists("llama.cpp"):
if IS_KAGGLE_ENVIRONMENT:
python_install = install_python_non_blocking(["protobuf"])
python_install.wait()
install_llama_cpp_blocking(use_cuda=False)
makefile = None
else:
git_clone = install_llama_cpp_clone_non_blocking()
python_install = install_python_non_blocking(["protobuf"])
git_clone.wait()
makefile = install_llama_cpp_make_non_blocking()
python_install.wait()
else:
makefile = None
for _ in range(3):
gc.collect()
# Use the provided save_directory for local saving
save_lora_to_custom_dir(self, tokenizer, save_directory)
model_type = self.config.model_type
output_file = os.path.join(save_directory, "ggml-adapter-model.bin")
print(f"Unsloth: Converting auto-saved LoRA adapters at {save_directory} to GGML format.")
print(f"The output file will be {output_file}")
command = f"python3 llama.cpp/convert-lora-to-ggml.py {save_directory} {output_file} llama"
try:
with subprocess.Popen(command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=1, universal_newlines=True) as sp:
for line in sp.stdout:
print(line, end="", flush=True)
for line in sp.stderr:
print(line, end="", flush=True)
sp.wait()
if sp.returncode != 0:
raise subprocess.CalledProcessError(sp.returncode, command)
except subprocess.CalledProcessError as e:
print(f"Error: Conversion failed with return code {e.returncode}")
return
print("Unsloth: Done.")
print(f"Unsloth: Conversion completed! Output file: {output_file}")
print("\nThis GGML making function was made by Maheswar. Ping him @Maheswar on the Unsloth Discord or on HuggingFace (@mahiatlinux) if you like this!")
def patch_saving_functions(model):
import inspect
import types
from typing import Callable, Optional, Union, List
# And now re add our saving methods!
if model.push_to_hub.__name__ == "unsloth_push_to_hub":
original_push_to_hub = model.original_push_to_hub
else:
original_push_to_hub = model.push_to_hub
pass
signature = str(inspect.signature(original_push_to_hub)).replace("NoneType", "None")
signature = signature[1:]
signature = re.sub("<function save at .+?>", "torch.save", signature)
docs = original_push_to_hub.__doc__.encode("utf-8").decode("utf-8")
push_to_hub_text = f'''def unsloth_push_to_hub(self, {signature}:
"""
{docs}
"""
arguments = dict(locals())
del arguments["self"]
if "tags" in arguments and arguments["tags"] is not None:
assert(isinstance(arguments["tags"], (list, tuple)))
arguments["tags"] = list(arguments["tags"]) + ["unsloth",]
elif "tags" in arguments:
arguments["tags"] = ["unsloth",]
elif hasattr(self, "add_model_tags"):
self.add_model_tags(["unsloth",])
if "commit_message" in arguments:
commit_message = arguments["commit_message"]
if commit_message is not None:
if not commit_message.endswith(" "): commit_message += " "
if "Unsloth" not in commit_message:
commit_message += "(Trained with Unsloth)"
else:
commit_message = "Upload model trained with Unsloth"
arguments["commit_message"] = commit_message
if "commit_description" in arguments:
commit_description = arguments["commit_description"]
if commit_description is not None:
if not commit_description.endswith(" "): commit_description += " "
if "Unsloth" not in commit_description:
commit_description += "(Trained with Unsloth 2x faster)"
else:
commit_description = "Upload model trained with Unsloth 2x faster"
arguments["commit_description"] = commit_description
# Update model tag
if hasattr(self, "config"):
_ = upload_to_huggingface(
self, arguments["repo_id"], arguments["token"],
"finetuned", "trl", file_location = None,
old_username = None, private = arguments["private"],
)
pass
try:
self.original_push_to_hub(**arguments)
except:
del arguments["tags"]
self.original_push_to_hub(**arguments)
pass
if hasattr(self, "config"):
print("Saved model to https://huggingface.co/" + arguments["repo_id"])
pass
'''
exec(push_to_hub_text, globals())
original_model = model
while True:
if original_model.push_to_hub.__name__ != "unsloth_push_to_hub":
original_model.original_push_to_hub = original_model.push_to_hub
original_model.push_to_hub = types.MethodType(unsloth_push_to_hub, original_model)
if hasattr(original_model, "add_model_tags"):
original_model.add_model_tags(["unsloth",])
pass
pass
if hasattr(original_model, "model"): original_model = original_model.model
else: break
pass
# Add saving methods to top level model
if hasattr(model, "config"):
# Counteract tokenizers
model.push_to_hub_merged = types.MethodType(unsloth_push_to_hub_merged, model)
model.save_pretrained_merged = types.MethodType(unsloth_save_pretrained_merged, model)
model.push_to_hub_gguf = types.MethodType(unsloth_push_to_hub_gguf, model)
model.save_pretrained_gguf = types.MethodType(unsloth_save_pretrained_gguf, model)
model.push_to_hub_ggml = types.MethodType(unsloth_convert_lora_to_ggml_and_push_to_hub, model)
model.save_pretrained_ggml = types.MethodType(unsloth_convert_lora_to_ggml_and_save_locally, model)
pass
return model
pass
build/lib/build/lib/build/lib/unsloth/tokenizer_utils.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from transformers import AutoTokenizer
from transformers.convert_slow_tokenizer import convert_slow_tokenizer
from transformers import PreTrainedTokenizerFast
import re
import os
from transformers.models.llama.modeling_llama import logger
from peft import PeftModelForCausalLM
import torch
import itertools
import collections
import numpy as np
import gc
import subprocess
__all__ = [
"load_correct_tokenizer",
"fix_sentencepiece_tokenizer",
"check_tokenizer",
"add_new_tokens",
"fix_sentencepiece_gguf",
]
IGNORED_TOKENIZER_CHECKING = frozenset((
"CodeLlamaTokenizerFast",
"CodeLlamaTokenizer",
))
IGNORED_TOKENIZER_NAMES = [
# "unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit",
# "unsloth/Mistral-Nemo-Instruct-2407",
# "mistralai/Mistral-Nemo-Instruct-2407",
# "unsloth/Mistral-Nemo-Base-2407-bnb-4bit",
# "unsloth/Mistral-Nemo-Base-2407",
# "mistralai/Mistral-Nemo-Base-2407",
]
IGNORED_TOKENIZER_NAMES = frozenset(
[x.lower() for x in IGNORED_TOKENIZER_NAMES]
)
# Check environments
keynames = "\n" + "\n".join(os.environ.keys())
IS_COLAB_ENVIRONMENT = "\nCOLAB_" in keynames
IS_KAGGLE_ENVIRONMENT = "\nKAGGLE_" in keynames
del keynames
def try_fix_tokenizer(tokenizer, prepend = True):
if hasattr(tokenizer, "_tokenizer"):
converted_tokenizer = tokenizer._tokenizer
else:
converted_tokenizer = convert_slow_tokenizer(tokenizer)
pass
tokenizer_string = converted_tokenizer.to_str()
# Llama does _apple. Sometimes this is wrong!!
prepend_text = '{"type":"Prepend","prepend":"▁"},'
if not prepend and prepend_text in tokenizer_string:
tokenizer_string = tokenizer_string.replace(prepend_text, "", 1)
pass
dir_names = dir(tokenizer)
# Get eos_token, bos_token etc
token_names = [x for x in dir_names if x.endswith("_token") and x.count("_") == 1]
for token_name in token_names:
token = getattr(tokenizer, token_name, None)
if token is None: continue
token_id = getattr(tokenizer, token_name + "_id", None)
# Locate the token's id mapping in the string
find_text = f'"id":{token_id},"content":"'
start = tokenizer_string.find(find_text) + len(find_text)
if start == -1: continue
end = tokenizer_string.find('",', start)
bad_token = tokenizer_string[start : end]
# Check if token is the actual same one - if not, edit it
if bad_token != token:
bad_text = f'{find_text}{bad_token}",'
good_text = f'{find_text}{token}",'
tokenizer_string = tokenizer_string.replace(bad_text, good_text, 1)
# And replace vocab section
bad_text = f'"{bad_token}":{token_id},'
good_text = f'"{token}":{token_id},'
tokenizer_string = tokenizer_string.replace(bad_text, good_text, 1)
pass
pass
fixed_tokenizer = converted_tokenizer.from_str(tokenizer_string)
return fixed_tokenizer
pass
def get_sorted_dict(dictionary):
sorted_keys = sorted(dictionary.values())
inverted_dictionary = { value : key for key, value in dictionary.items() }
sorted_dictionary = {}
for key in sorted_keys:
value = inverted_dictionary[key]
sorted_dictionary[value] = key
return sorted_dictionary
pass
def convert_to_fast_tokenizer(
slow_tokenizer,
temporary_location = "_unsloth_sentencepiece_temp",
):
is_fast = getattr(slow_tokenizer, "is_fast", False)
if is_fast: return slow_tokenizer
try:
tokenizer_name = slow_tokenizer.__class__.__name__
lowered_tokenizer_name = tokenizer_name.lower()
if lowered_tokenizer_name.endswith("tokenizer"):
class_name = lowered_tokenizer_name[:-len("tokenizer")]
FastTokenizer = eval(
f'__import__(f"transformers.models.{class_name}").{tokenizer_name}Fast'
)
else:
FastTokenizer = PreTrainedTokenizerFast
except:
FastTokenizer = PreTrainedTokenizerFast
pass
# Get all arguments (bos_token, etc)
docs = FastTokenizer.__doc__
docs = docs[docs.find("Args:"):]
args = re.findall(r"\n[\s]+([^\s]{1,}) \(", docs, flags = re.MULTILINE)
args = [x for x in args if not x.endswith("_file")]
# Also some missing maybe!
docs = PreTrainedTokenizerFast.__doc__
docs = docs[docs.find("Args:"):]
args2 = re.findall(r"\n[\s]+([^\s]{1,}) \(", docs, flags = re.MULTILINE)
args2 = [x for x in args2 if not x.endswith("_file")]
args = list(set(args + args2))
kwargs = {}
for arg in args: kwargs[arg] = getattr(slow_tokenizer, arg, None)
kwargs["tokenizer_object"] = try_fix_tokenizer(slow_tokenizer, prepend = True)
fast_tokenizer = FastTokenizer( **kwargs )
# Check if they're similar!
sorted_slow_tokenizer = get_sorted_dict(slow_tokenizer.get_vocab())
sorted_fast_tokenizer = get_sorted_dict(fast_tokenizer.get_vocab())
check_vocab = (sorted_slow_tokenizer == sorted_fast_tokenizer)
check_special = (slow_tokenizer.all_special_tokens == fast_tokenizer.all_special_tokens)
# Failure so return slow_tokenizer
if not check_vocab or not check_special: return slow_tokenizer
# Now confirm if they match
if not assert_same_tokenization(slow_tokenizer, fast_tokenizer):
# Maybe remove prepending of __apple?
kwargs["tokenizer_object"] = try_fix_tokenizer(slow_tokenizer, prepend = False)
fast_tokenizer = FastTokenizer( **kwargs )
if not assert_same_tokenization(slow_tokenizer, fast_tokenizer):
# Failure :(
return slow_tokenizer
pass
pass
# Also tokenizer.model is missing!
name = slow_tokenizer.name_or_path.replace("/", "_")
if not os.path.exists(temporary_location):
os.makedirs(temporary_location)
pass
new_location = f"{temporary_location}/{name}"
slow_tokenizer.save_pretrained(new_location)
fast_tokenizer.save_pretrained(new_location)
# Now load it!
fast_tokenizer = AutoTokenizer.from_pretrained(new_location)
if assert_same_tokenization(slow_tokenizer, fast_tokenizer):
return fast_tokenizer
return slow_tokenizer
pass
# Check Mistral chat template without BOS / EOS
mistral_template = \
"{% if messages[0]['role'] == 'system' %}"\
"{% if messages[1]['role'] == 'user' %}"\
"{{ '[INST] ' + messages[0]['content'] + ' ' + messages[1]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[2:] %}"\
"{% else %}"\
"{{ '[INST] ' + messages[0]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% endif %}"\
"{% else %}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '[INST] ' + message['content'] + ' [/INST]' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ message['content'] }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"
pass
# Check Llama chat template without BOS / EOS
llama_template = \
"{% if messages[0]['role'] == 'system' %}"\
"{% if messages[1]['role'] == 'user' %}"\
"{{ '[INST] <<SYS>>\n' + messages[0]['content'] + '\n<</SYS>>\n\n' + messages[1]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[2:] %}"\
"{% else %}"\
"{{ '[INST] ' + messages[0]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% endif %}"\
"{% else %}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '[INST] ' + message['content'].strip() + ' [/INST]' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ ' ' + message['content'].strip() + ' ' }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"
pass
def assert_same_tokenization(slow_tokenizer, fast_tokenizer):
# Get eos_token, bos_token etc
dir_names = dir(slow_tokenizer)
special_tokens = list(filter(None, (
getattr(slow_tokenizer, x) for x in dir_names
if x.endswith("_token") and x.count("_") == 1
)))
all_special_tokens = list(set(special_tokens + slow_tokenizer.all_special_tokens))
# Check if chat template is enabled!
check_chat_template1 = True
check_chat_template2 = True
check_chat_template3 = True
"""
Weirdly Mistral tokenizers are actually correct??
Ie below will actually load mistral v1 and v3 incorrectly!
slow_chat_template = getattr(slow_tokenizer, "chat_template", None)
fast_chat_template = getattr(fast_tokenizer, "chat_template", None)
messages = [
{"role": "user", "content": " What is 2+2? "},
{"role": "assistant", "content": " It's 4. "},
]
# Check the tokenizer's own chat template
if slow_chat_template is not None and fast_chat_template is not None:
check_chat_template1 = \
slow_tokenizer.apply_chat_template(messages) == \
fast_tokenizer.apply_chat_template(messages)
pass
# Check Mistral chat template without BOS / EOS
slow_tokenizer.chat_template = mistral_template
fast_tokenizer.chat_template = mistral_template
check_chat_template2 = \
slow_tokenizer.apply_chat_template(messages) == \
fast_tokenizer.apply_chat_template(messages)
pass
# Check Llama chat template without BOS / EOS
slow_tokenizer.chat_template = llama_template
fast_tokenizer.chat_template = llama_template
check_chat_template3 = \
slow_tokenizer.apply_chat_template(messages) == \
fast_tokenizer.apply_chat_template(messages)
pass
# Combine them all and revert chat templates
slow_tokenizer.chat_template = slow_chat_template
fast_tokenizer.chat_template = fast_chat_template
"""
check_chat_template = check_chat_template1 and check_chat_template2 and check_chat_template3
# Try special tokens
try:
string = "\n".join(all_special_tokens) + \
"A quick brown fox jumps over the lazy dog!!\n\nHi</s>\n\n" + \
"".join(all_special_tokens)
check_special_tokens = \
slow_tokenizer(string).input_ids == \
fast_tokenizer(string).input_ids
return check_chat_template and check_special_tokens
except:
# For eg see https://github.com/unslothai/unsloth/issues/292
# Sometimes tokenizer has weird tokens, causing a combined tokenization to fail.
# [TODO] We temporarily disable this for CodeLlama tokenizers
if slow_tokenizer.__repr__().split("(", 1)[0] in IGNORED_TOKENIZER_CHECKING:
return check_chat_template
else:
return False
pass
pass
def fix_sentencepiece_tokenizer(
old_tokenizer,
new_tokenizer,
token_mapping,
temporary_location = "_unsloth_sentencepiece_temp",
):
# From https://github.com/google/sentencepiece/issues/121
# We need to manually edit the sentencepiece tokenizer!
from transformers.utils import sentencepiece_model_pb2
if not os.path.exists(temporary_location):
os.makedirs(temporary_location)
pass
# Check if tokenizer.model exists
if not os.path.isfile(f"{temporary_location}/tokenizer.model"):
return new_tokenizer
pass
# First save the old tokenizer
old_tokenizer.save_pretrained(temporary_location)
tokenizer_file = sentencepiece_model_pb2.ModelProto()
tokenizer_file.ParseFromString(open(f"{temporary_location}/tokenizer.model", "rb").read())
# Now save the new tokenizer
new_tokenizer.save_pretrained(temporary_location)
# Now correct the old tokenizer's .model file
for old_token, new_token in token_mapping.items():
ids = old_tokenizer([old_token], add_special_tokens = False).input_ids
ids = ids[0]
if (len(ids) != 1):
# Skip this token!
print(f"Skip mapping {old_token} to {new_token} since {new_token} is already in the tokenizer!")
continue
pass
ids = ids[0]
# [TODO] Hack for Starling - try except
try:
tokenizer_piece = tokenizer_file.pieces[ids]
except:
continue
assert(tokenizer_piece.piece == old_token)
tokenizer_piece.piece = new_token
pass
# And now write it
with open(f"{temporary_location}/tokenizer.model", "wb") as file:
file.write(tokenizer_file.SerializeToString())
pass
# And load it!
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(
temporary_location,
eos_token = new_tokenizer.eos_token,
pad_token = new_tokenizer.pad_token,
)
return tokenizer
pass
def fix_sentencepiece_gguf(saved_location):
"""
Fixes sentencepiece tokenizers which did not extend the vocabulary with
user defined tokens.
Inspiration from https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py
"""
from copy import deepcopy
from transformers.utils import sentencepiece_model_pb2
import json
from enum import IntEnum
class SentencePieceTokenTypes(IntEnum):
NORMAL = 1
UNKNOWN = 2
CONTROL = 3
USER_DEFINED = 4
UNUSED = 5
BYTE = 6
pass
# Load tokenizer.model
tokenizer_file = sentencepiece_model_pb2.ModelProto()
if not os.path.isfile(f"{saved_location}/tokenizer.model"): return
tokenizer_file.ParseFromString(open(f"{saved_location}/tokenizer.model", "rb").read())
sentence_piece_size = len(tokenizer_file.pieces)
# Load added_tokens_json
if not os.path.isfile(f"{saved_location}/added_tokens.json"): return
with open(f"{saved_location}/added_tokens.json", "r", encoding = "utf-8") as file:
added_tokens_json = json.load(file)
pass
if len(added_tokens_json) == 0: return
added_tokens_json = dict(sorted(added_tokens_json.items(), key = lambda item: item[1]))
new_size = sentence_piece_size + len(added_tokens_json)
# Confirm added_tokens_json is correct
added_tokens_ids = np.array(list(added_tokens_json.values()))
diff = np.diff(added_tokens_ids)
if (diff.min() != 1 or diff.max() != 1): return
if (added_tokens_ids.min() != sentence_piece_size): return
# Edit sentence piece tokens with added_tokens_json
logger.warning(
f"Unsloth: Extending {saved_location}/tokenizer.model with added_tokens.json.\n"\
f"Originally tokenizer.model is of size ({sentence_piece_size}).\n"\
f"But we need to extend to sentencepiece vocab size ({new_size})."
)
new_tokens = deepcopy(tokenizer_file.pieces[-len(added_tokens_ids):])
for new_token, added_token in zip(new_tokens, added_tokens_json.keys()):
new_token.piece = added_token.encode("utf-8")
new_token.score = -1000.0
new_token.type = SentencePieceTokenTypes.USER_DEFINED
pass
tokenizer_file.pieces.extend(new_tokens)
with open(f"{saved_location}/tokenizer.model", "wb") as file:
file.write(tokenizer_file.SerializeToString())
pass
# Add padding tokens
# actual_vocab_size = model.config.vocab_size
# padding = actual_vocab_size - len(tokenizer_file.pieces)
return
pass
def load_correct_tokenizer(
tokenizer_name,
model_max_length = None,
padding_side = "right",
token = None,
trust_remote_code = False,
cache_dir = "huggingface_tokenizers_cache",
):
if IS_COLAB_ENVIRONMENT or IS_KAGGLE_ENVIRONMENT:
cache_dir = cache_dir
else:
cache_dir = None
pass
# Try loading the slow tokenizer. If it fails, then try Fast only
# Mainly to solve Deepseek models with no tokenizer.model file
slow_tokenizer = None
try:
slow_tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name,
model_max_length = model_max_length,
padding_side = padding_side,
token = token,
trust_remote_code = trust_remote_code,
# Cannot just use use_fast = False as per https://twitter.com/danielhanchen/status/1789659394302718373
use_fast = False,
legacy = False,
from_slow = True,
cache_dir = cache_dir,
)
except:
pass
# print(
# f"Unsloth: {tokenizer_name} has no tokenizer.model file.\n"\
# "Just informing you about this - this is not a critical error."
# )
pass
fast_tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name,
model_max_length = model_max_length,
padding_side = padding_side,
token = token,
trust_remote_code = trust_remote_code,
cache_dir = cache_dir,
)
if tokenizer_name in IGNORED_TOKENIZER_NAMES:
return fast_tokenizer
elif slow_tokenizer is not None:
if hasattr(fast_tokenizer, "add_bos_token") and hasattr(slow_tokenizer, "add_bos_token"):
fast_tokenizer.add_bos_token = slow_tokenizer.add_bos_token
if hasattr(fast_tokenizer, "add_eos_token") and hasattr(slow_tokenizer, "add_eos_token"):
fast_tokenizer.add_eos_token = slow_tokenizer.add_eos_token
# Confirm if slow and fast are equivalent!
if assert_same_tokenization(slow_tokenizer, fast_tokenizer):
return fast_tokenizer
else:
logger.warning(f"Unsloth: Will load {tokenizer_name} as a legacy tokenizer.")
return convert_to_fast_tokenizer(slow_tokenizer)
pass
else:
return fast_tokenizer
pass
pass
def check_tokenizer(
model,
tokenizer,
model_name = "unsloth/llama-2-7b-bnb-4bit",
model_max_length = 4096,
padding_side = "right",
token = None,
_reload = True,
):
# Checks tokenizer for out of bounds ids.
# Mainly a fix for https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha
# where <sep> had token id=32002.
# See https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha/discussions/25
# Seems like the Fast tokenizer in Rust breaks things!
# We ignore some of them!
if tokenizer.__repr__().split("(", 1)[0] in IGNORED_TOKENIZER_CHECKING:
return tokenizer
pass
max_embedding_size = model.model.embed_tokens.weight.shape[0]
added_tokens_fast = tokenizer.added_tokens_decoder
added_tokens_fast = {index : str(value) for index, value in added_tokens_fast.items()}
sorted_keys = sorted(added_tokens_fast)
added_tokens_fast = {key : added_tokens_fast[key] for key in sorted_keys}
for j, index in enumerate(added_tokens_fast.keys()):
if index >= max_embedding_size:
bad_indices = list(added_tokens_fast.keys ())[j:]
bad_tokens = list(added_tokens_fast.values())[j:]
if not _reload:
# Try removing the token
added_tokens = [str(x) for x in tokenizer.added_tokens_decoder.values()]
special_tokens = tokenizer.special_tokens_map
import itertools
special_tokens = frozenset(
itertools.chain.from_iterable(
[x] if type(x) is str else x for x in special_tokens.values()
)
)
can_be_removed1 = [x for x in bad_tokens if x not in special_tokens]
can_be_removed2 = [x for x in can_be_removed1 if x in tokenizer._added_tokens_encoder.keys()]
# Check of extra tokens can in fact we removed!
can_be_removed = \
(len(can_be_removed1) == len(bad_tokens)) and \
(len(can_be_removed2) == len(bad_tokens))
# Check if sep_token or other generic types
remove_generic = False
try_mapper = []
if not can_be_removed:
names = dir(tokenizer)
names = (x for x in names if x.endswith("_token") and x.count("_") == 1)
generic_tokens = [(x, getattr(tokenizer, x, None)) for x in names]
try_removal = []
for token in bad_tokens:
for (name_token, check_token) in generic_tokens:
if check_token == token:
try_removal.append(token)
try_mapper.append(name_token)
pass
pass
pass
# Recheck!
can_be_removed = (len(try_removal) == len(bad_tokens))
if can_be_removed: remove_generic = True
can_be_removed1 = bad_tokens
pass
if can_be_removed:
# Yes it can be fixed!
for j, bad_token in enumerate(can_be_removed1):
remove_id = tokenizer._added_tokens_encoder[bad_token]
del tokenizer._added_tokens_decoder[remove_id]
del tokenizer._added_tokens_encoder[bad_token]
if remove_generic and (try_removal[j] == bad_token):
# Remove sep token for example
setattr(tokenizer, try_mapper[j], None)
setattr(tokenizer, try_mapper[j] + "_id", None)
pass
pass
# Confirm 1 more time!
if max(tokenizer.added_tokens_decoder.keys()) < max_embedding_size:
logger.warning_once(
f"Unsloth loaded a broken tokenizer `{model_name}`, but managed to repair it!\n"\
f"Tokens {bad_tokens} with ids {bad_indices} exceeds the max vocab size of {max_embedding_size}.\n"\
"We removed these bad tokens. If you think this is incorrect, fix your tokenizer first."
)
return convert_to_fast_tokenizer(tokenizer)
pass
pass
# :( Failure
raise RuntimeError(
f"Unsloth tried to load `{model_name}`, but cannot succeed.\n"\
f"Tokens {bad_tokens} with ids {bad_indices} exceeds the max vocab size of {max_embedding_size}.\n"\
f"Fix your tokenizer since it'll perform out of bounds memory accesses."
)
pass
if IS_COLAB_ENVIRONMENT or IS_KAGGLE_ENVIRONMENT:
cache_dir = "huggingface_tokenizers_cache"
else:
cache_dir = None
pass
# Sometimes slow tokenizer does not work like Deepseek
try:
# Try slow tokenizer which can fix things!
tokenizer = AutoTokenizer.from_pretrained(
model_name,
model_max_length = model_max_length,
padding_side = padding_side,
token = token,
# Cannot just use use_fast = False as per https://twitter.com/danielhanchen/status/1789659394302718373
use_fast = False,
legacy = False,
from_slow = True,
cache_dir = cache_dir,
)
return check_tokenizer(
model = model,
tokenizer = tokenizer,
model_name = model_name,
model_max_length = model_max_length,
padding_side = padding_side,
token = token,
_reload = False,
)
break
except:
# Tokenizer has out of bounds issues and we can't
# load the slow tokenizer version :(
logger.warning_once(
"Unsloth: Tokenizer is most likely buggy, and Unsloth failed to repair it.\n"\
"It will still work, but beware of out of bounds memory accesses.\n"\
"Please file an issue on the model owner's repo about this issue."
)
return tokenizer
pass
pass
pass
return convert_to_fast_tokenizer(tokenizer)
pass
@torch.inference_mode
def fix_untrained_tokens(model, tokenizer, train_dataset, eps = 1e-16):
"""
Llama-3 for eg has untrained vectors in the base model.
These include <|eot_id|>, <|start_header_id|>, <|end_header_id|>
We reset them to the mean of the rest of the tokens
"""
embedding_matrix = model.get_input_embeddings ().weight
lm_head_matrix = model.get_output_embeddings().weight
# Ignore some model checks for now
if model.config._name_or_path in IGNORED_TOKENIZER_NAMES:
return
pass
# Get untrained tokens
indicator_untrained1 = torch.amax(embedding_matrix, axis = 1) <= eps
# Check lm_head as well
indicator_untrained2 = torch.amax(lm_head_matrix, axis = 1) <= eps
# Combine both checks
indicator_untrained = indicator_untrained1 & indicator_untrained2
where_untrained = torch.where(indicator_untrained)[0]
n_untrained = where_untrained.shape[0]
n_trained = embedding_matrix.shape[0] - n_untrained
# Get set and actual tokens
where_untrained = where_untrained.tolist()
if len(where_untrained) == 0: return
# Remove untrained indices where it's longer
where_untrained_set = frozenset(where_untrained)
actual_bad_tokens = tokenizer.convert_ids_to_tokens(where_untrained)
# Remove None items in actual_bad_tokens
actual_bad_tokens = [x for x in actual_bad_tokens if x is not None]
# Check if tokenizer and training datasets have bad tokens
if_bad_first = False
if_bad_second = False
# Check tokenizer's chat template for any untrained tokens
chat_template = getattr(tokenizer, "chat_template", None)
if chat_template is not None:
if_bad_first = any(x in chat_template for x in actual_bad_tokens)
pass
# Check the first 250, last 250 input_ids
size_dataset = len(train_dataset)
size = min(size_dataset, 250)
for j in range(size):
input_ids = train_dataset[j]
if "input_ids" in input_ids:
input_ids = input_ids["input_ids"]
if_bad = any(item in where_untrained_set for item in input_ids)
if if_bad:
if_bad_second = True
break
pass
pass
pass
# Check last 250
if not if_bad_second:
left = max(size_dataset-250, 0)
for j in range(left, size_dataset):
input_ids = train_dataset[j]
if "input_ids" in input_ids:
input_ids = input_ids["input_ids"]
if_bad = any(item in where_untrained_set for item in input_ids)
if if_bad:
if_bad_second = True
break
pass
pass
pass
pass
# Check if bad tokens exists!
if not if_bad_first and not if_bad_second: return
# Check if lm_head / embed_token are trainable!
bad_not_trainable = False
if not embedding_matrix.requires_grad: bad_not_trainable = True
if not lm_head_matrix .requires_grad: bad_not_trainable = True
if bad_not_trainable:
raise ValueError(
'Unsloth: Untrained tokens found, but embed_tokens & lm_head not trainable, causing NaNs. '\
'Restart then add `embed_tokens` & `lm_head` to '\
'`FastLanguageModel.get_peft_model(target_modules = [..., "embed_tokens", "lm_head",]). `'\
'Are you using the `base` model? Instead, use the `instruct` version to silence this warning.',
)
pass
# Count all the possible bad tokens
final_counts = np.zeros(max(len(tokenizer), embedding_matrix.shape[0]), dtype = np.int64)
def mapping(examples):
input_ids = examples["input_ids"]
counter = np.fromiter(itertools.chain.from_iterable(input_ids), dtype = np.int32)
np.add.at(final_counts, counter, 1)
pass
train_dataset.map(mapping, batched = True, desc = "Counting untrained tokens")
# Get sum of all items
sum_embedding = torch.sum(embedding_matrix, dtype = torch.float32, axis = 0)
sum_lm_head = torch.sum(lm_head_matrix, dtype = torch.float32, axis = 0)
# Remove bad tokens
sum_embedding -= torch.sum(embedding_matrix[where_untrained], dtype = torch.float32, axis = 0)
sum_lm_head -= torch.sum(lm_head_matrix [where_untrained], dtype = torch.float32, axis = 0)
# Find correct average by dividing by sum of trained tokens
mean_embedding = (sum_embedding / n_trained)
mean_lm_head = (sum_lm_head / n_trained)
# Scale each to be equal to 1/max_frequency. Also set some to 0 if none seen
scaling = final_counts[where_untrained] / max(final_counts.max(), 1)
scaling = torch.tensor(scaling, device = mean_embedding.device).unsqueeze(1)
mean_embedding = mean_embedding.repeat((n_untrained, 1,)) * scaling
mean_lm_head = mean_lm_head .repeat((n_untrained, 1,)) * scaling
where_null = scaling.ravel() == 0
mean_embedding[where_null] = 0
mean_lm_head [where_null] = 0
# Set them to the mean
logger.warning(
"Unsloth: Setting embed_tokens & lm_head untrained tokens to "\
"mean(trained) to counteract NaNs during training."
)
embedding_matrix[where_untrained] = mean_embedding.to(embedding_matrix.dtype)
lm_head_matrix [where_untrained] = mean_lm_head .to(lm_head_matrix .dtype)
# Clean up
for _ in range(3):
gc.collect()
torch.cuda.empty_cache()
pass
return
pass
@torch.inference_mode
def mean_of_trained_tokens(model, eps = 1e-16):
"""
Llama-3 for eg has untrained vectors in the base model.
These include <|eot_id|>, <|start_header_id|>, <|end_header_id|>
We reset them to the mean of the rest of the tokens
"""
embedding_matrix = model.get_input_embeddings ().weight.clone()
lm_head_matrix = model.get_output_embeddings().weight.clone()
# Get untrained tokens
indicator_untrained = torch.amax(embedding_matrix, axis = 1) <= eps
where_untrained = torch.where(indicator_untrained)[0]
n_untrained = where_untrained.shape[0]
n_trained = embedding_matrix.shape[0] - n_untrained
# if n_untrained != 0:
# print(
# f"Unsloth: Not an error, but your model has {n_untrained} untrained tokens.\n"\
# "We shall set them to the mean of the other trained tokens."
# )
# pass
# Get sum of all items
sum_embedding = torch.sum(embedding_matrix, dtype = torch.float32, axis = 0)
sum_lm_head = torch.sum(lm_head_matrix, dtype = torch.float32, axis = 0)
# Remove bad tokens
sum_embedding -= torch.sum(embedding_matrix[where_untrained], dtype = torch.float32, axis = 0)
sum_lm_head -= torch.sum(lm_head_matrix [where_untrained], dtype = torch.float32, axis = 0)
# Find correct average by dividing by sum of trained tokens
mean_embedding = (sum_embedding / n_trained)
mean_lm_head = (sum_lm_head / n_trained)
return mean_embedding, mean_lm_head
pass
@torch.inference_mode
def add_new_tokens(
model,
tokenizer,
new_tokens = [],
method = "mean",
interpolation = 0.5,
):
"""
Smartly resizes the tokenizer and adds new tokens to the model.
We also disregard untrained tokens by removing them from the mean calculation.
"""
assert(isinstance(new_tokens, (list, tuple)))
assert(len(new_tokens) > 0)
assert(method == "mean" or method == "interpolation")
assert(interpolation >= 0 and interpolation <= 1)
# Check if tokens already exist
overlapping_tokens = set(new_tokens) & set(tokenizer.vocab.keys())
if len(overlapping_tokens) != 0:
print(
f"Unsloth: You're adding new_tokens = {new_tokens}\n"\
f"There are tokens which are overlapping = {list(overlapping_tokens)}\n"\
f"We shall safely ignore these overlapping tokens."
)
new_tokens = [x for x in new_tokens if x not in overlapping_tokens]
pass
# Get mean of trained tokens
# mean_embedding, mean_lm_head = fix_untrained_tokens(model)
# Weirdly be careful reserved tokens can pop out
mean_embedding, mean_lm_head = mean_of_trained_tokens(model)
mean_embedding = mean_embedding.to(torch.float32)
mean_lm_head = mean_lm_head .to(torch.float32)
# Add tokens!
old_length = len(tokenizer)
tokenizer.add_tokens(new_tokens)
model.resize_token_embeddings(len(tokenizer))
# If we use interpolation, we interpolate between the mean embeddings and
# the Word2Vec sum of the other vectors
embedding_matrix = model.get_input_embeddings ().weight
lm_head_matrix = model.get_output_embeddings().weight
if method == "interpolation":
print(
"Unsloth: You are using interpolation to add new tokens.\n"\
f"We shall set new tokens = mean(embeddings)*{1-interpolation} + mean(new_tokens)*{interpolation}"
)
for j, token in enumerate(new_tokens):
input_ids = tokenizer(token, add_special_tokens = False).input_ids
mean_embedding_token = embedding_matrix[input_ids].mean(axis = 0, dtype = torch.float32)
mean_lm_head_token = lm_head_matrix [input_ids].mean(axis = 0, dtype = torch.float32)
# Interpolate
mean_embedding_token = mean_embedding*(1-interpolation) + mean_embedding_token*interpolation
mean_lm_head_token = mean_lm_head *(1-interpolation) + mean_lm_head_token *interpolation
# Set the new vector
embedding_matrix[old_length+j] = mean_embedding_token
lm_head_matrix [old_length+j] = mean_lm_head_token
pass
else:
# Now set the new tokens to the mean!
embedding_matrix[old_length:] = mean_embedding
lm_head_matrix [old_length:] = mean_lm_head
pass
# We set a flag to say we need to train embeddings
internal_model = model
while hasattr(internal_model, "model"):
internal_model._need_to_train_embeddings = True
internal_model = internal_model.model
pass
internal_model._need_to_train_embeddings = True
return
pass
def check_nvidia():
# Unsloth doesn't work yet on AMD devices - we're working on it!
output = np.array([0,])
try:
output = subprocess.check_output("nvidia-smi --query-gpu=memory.used --format=csv", shell = True)
output = re.findall(rb'([\d]{1,})[\s]{1,}M', output)
output = np.array([int(x.decode('utf-8'))/1024 for x in output])
except:
if not torch.cuda.is_available():
raise RuntimeError("Unsloth: We do not support AMD / Intel machines yet - it is a work in progress!")
return output
pass
PRE_CHECK = check_nvidia()
from inspect import getsource
import trl.trainer.sft_trainer
from trl.trainer.sft_trainer import *
from transformers.trainer import *
def patch_sft_trainer_tokenizer():
"""
Patches the trainer with changes
"""
for function_name, replacer in (
("_prepare_non_packed_dataloader", "def tokenize(element):",),
# ("_prepare_packed_dataloader", "if dataset_text_field is not None",),
):
function = getsource(eval(f"trl.trainer.sft_trainer.SFTTrainer.{function_name}"))
where = function.find("def")
function = function.split("\n")
function = "\n".join(x[where:] for x in function)
check_text = \
"\n"\
"test_text = dataset[0][dataset_text_field] if (formatting_func is None or not use_formatting_func) else formatting_func(dataset[0])[0]\n"\
"chat_template = getattr(tokenizer, 'chat_template', None)\n"\
"chat_template = '' if chat_template is None else chat_template\n"\
"has_bos_token_already = (test_text.startswith(tokenizer.bos_token) or tokenizer.bos_token in chat_template) "\
"if getattr(tokenizer, 'bos_token', None) is not None else False\n"\
"add_special_tokens = False if has_bos_token_already else add_special_tokens\n\n"
check_text = check_text.split("\n")
check_text = "\n".join(" "*where + x for x in check_text)
function = function.replace(replacer, check_text + replacer)
exec(function, globals())
exec(f"trl.trainer.sft_trainer.SFTTrainer.{function_name} = {function_name}", globals())
pass
# Patch train with fix_untrained_tokens
function_name, replacer = "train", "if resume_from_checkpoint is False:"
function = getsource(eval(f"trl.trainer.sft_trainer.SFTTrainer.{function_name}"))
where = function.find("def")
function = function.split("\n")
function = "\n".join(x[where:] for x in function)
check_text = \
"\n"\
"if self._inner_training_loop.__name__ != '_fast_inner_training_loop':\n"\
" raise RuntimeError(\n"\
" 'Please do not edit specific areas of the Unsloth codebase or you will get CUDA segfaults.'\n"\
" )\n"\
"pass\n"\
"import subprocess, re, gc, numpy as np\n"\
"a = np.array([0,])\n"\
"try:\n"\
" a = subprocess.check_output('nvidia-smi --query-gpu=memory.used --format=csv', shell = True)\n"\
" a = re.findall(rb'([\\d]{1,})[\\s]{1,}M', a)\n"\
" a = np.array([int(x.decode('utf-8'))/1024 for x in a])\n"\
"except:\n"\
" if not torch.cuda.is_available():\n"\
" raise RuntimeError('Unsloth: We do not support AMD / Intel machines yet - it is a work in progress!')\n"\
"if ((a - PRE_CHECK) >= 1).sum() > 1:\n"\
" raise RuntimeError('Unsloth currently does not support multi GPU setups - but we are working on it!')\n"\
"for _ in range(3):\n"\
" gc.collect()\n"\
" torch.cuda.empty_cache()\n"\
"pass\n"\
"\n"\
"fix_untrained_tokens(self.model, self.tokenizer, self.train_dataset, eps = 1e-16)\n\n"
check_text = check_text.split("\n")
check_text = "\n".join(" "*where + x for x in check_text)
function = function.replace(replacer, check_text + replacer)
exec(function, globals())
exec(f"trl.trainer.sft_trainer.SFTTrainer.{function_name} = {function_name}", globals())
pass
patch_sft_trainer_tokenizer()
build/lib/build/lib/build/lib/unsloth/trainer.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass, field
from typing import Optional
from transformers import TrainingArguments
from trl import SFTTrainer
from . import is_bfloat16_supported
__all__ = [
"UnslothTrainingArguments",
"UnslothTrainer",
]
@dataclass
class UnslothTrainingArguments(TrainingArguments):
embedding_learning_rate : Optional[float] = field(
default = None,
metadata = {"help" : "Different learning rates for embeddings and lm_head."}
)
pass
def _create_unsloth_optimizer(
model,
optimizer_cls,
optimizer_kwargs,
embedding_lr = 5e-5,
):
lr = optimizer_kwargs["lr"]
weight_decay = optimizer_kwargs.get("weight_decay", 0.0)
param_groups = \
{
"non_embeddings" : {},
"embeddings" : {},
}
for name, param in model.named_parameters():
if not param.requires_grad: continue
if name.endswith("modules_to_save.default.weight"):
partial_name = name[:-len(".modules_to_save.default.weight")]
partial_name = partial_name[partial_name.rfind(".")+1:]
print(f"Unsloth: Setting lr = {embedding_lr:.2e} instead of {lr:.2e} for {partial_name}.")
param_groups["embeddings"] [name] = param
else:
param_groups["non_embeddings"][name] = param
pass
pass
optimizer_grouped_parameters = [
{
"params" : list(param_groups["non_embeddings"].values()),
"weight_decay" : weight_decay,
"lr" : lr,
},
{
"params" : list(param_groups["embeddings"].values()),
"weight_decay" : weight_decay,
"lr" : embedding_lr,
},
]
optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
return optimizer
pass
class UnslothTrainer(SFTTrainer):
def create_optimizer(self):
embedding_learning_rate = getattr(self.args, "embedding_learning_rate", None)
if embedding_learning_rate is None: return super().create_optimizer()
if self.optimizer is None:
optimizer_cls, optimizer_kwargs = SFTTrainer.get_optimizer_cls_and_kwargs(self.args)
self.optimizer = _create_unsloth_optimizer(
self.model,
optimizer_cls,
optimizer_kwargs,
embedding_learning_rate,
)
pass
return self.optimizer
pass
pass
build/lib/build/lib/unsloth/__init__.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import warnings
import importlib
import sys
from packaging.version import Version
# # Define a list of modules to check
# MODULES_TO_CHECK = ["bitsandbytes"]
# # Check if any of the modules in the list have been imported
# for module in MODULES_TO_CHECK:
# if module in sys.modules:
# raise ImportError(f"Unsloth: Please import Unsloth before {module}.")
# pass
# pass
# Unsloth currently does not work on multi GPU setups - sadly we are a 2 brother team so
# enabling it will require much more work, so we have to prioritize. Please understand!
# We do have a beta version, which you can contact us about!
# Thank you for your understanding and we appreciate it immensely!
if "CUDA_VISIBLE_DEVICES" in os.environ:
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
devices = os.environ["CUDA_VISIBLE_DEVICES"]
# Check if there are multiple cuda devices set in env
if not devices.isdigit():
first_id = devices.split(",")[0]
warnings.warn(
f"Unsloth: 'CUDA_VISIBLE_DEVICES' is currently {devices} \n"\
"Unsloth currently does not support multi GPU setups - but we are working on it!\n"\
"Multiple CUDA devices detected but we require a single device.\n"\
f"We will override CUDA_VISIBLE_DEVICES to first device: {first_id}."
)
os.environ["CUDA_VISIBLE_DEVICES"] = str(first_id)
else:
# warnings.warn("Unsloth: 'CUDA_VISIBLE_DEVICES' is not set. We shall set it ourselves.")
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
pass
# Reduce VRAM usage by reducing fragmentation
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
try:
import torch
except:
raise ImportError("Pytorch is not installed. Go to https://pytorch.org/.\n"\
"We have some installation instructions on our Github page.")
pass
# Hugging Face Hub faster downloads (only enable during Colab and Kaggle sessions)
keynames = "\n" + "\n".join(os.environ.keys())
if "\nCOLAB_" in keynames or "\nKAGGLE_" in keynames:
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
pass
# We support Pytorch 2
# Fixes https://github.com/unslothai/unsloth/issues/38
torch_version = torch.__version__.split(".")
major_torch, minor_torch = torch_version[0], torch_version[1]
major_torch, minor_torch = int(major_torch), int(minor_torch)
if (major_torch < 2):
raise ImportError("Unsloth only supports Pytorch 2 for now. Please update your Pytorch to 2.1.\n"\
"We have some installation instructions on our Github page.")
elif (major_torch == 2) and (minor_torch < 2):
# Disable expandable_segments
del os.environ["PYTORCH_CUDA_ALLOC_CONF"]
pass
# Torch 2.5 has including_emulation
major_version, minor_version = torch.cuda.get_device_capability()
SUPPORTS_BFLOAT16 = (major_version >= 8)
if (major_torch == 2) and (minor_torch >= 5):
old_is_bf16_supported = torch.cuda.is_bf16_supported
def is_bf16_supported(including_emulation = False):
return old_is_bf16_supported(including_emulation)
torch.cuda.is_bf16_supported = is_bf16_supported
else:
def is_bf16_supported(): return SUPPORTS_BFLOAT16
torch.cuda.is_bf16_supported = is_bf16_supported
pass
# Try loading bitsandbytes and triton
import bitsandbytes as bnb
import triton
libcuda_dirs = lambda: None
if Version(triton.__version__) >= Version("3.0.0"):
try: from triton.backends.nvidia.driver import libcuda_dirs
except: pass
else: from triton.common.build import libcuda_dirs
import os
import re
import numpy as np
import subprocess
try:
cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
libcuda_dirs()
except:
warnings.warn(
"Unsloth: Running `ldconfig /usr/lib64-nvidia` to link CUDA."\
)
if os.path.exists("/usr/lib64-nvidia"):
os.system("ldconfig /usr/lib64-nvidia")
elif os.path.exists("/usr/local"):
# Sometimes bitsandbytes cannot be linked properly in Runpod for example
possible_cudas = subprocess.check_output(["ls", "-al", "/usr/local"]).decode("utf-8").split("\n")
find_cuda = re.compile(r"[\s](cuda\-[\d\.]{2,})$")
possible_cudas = [find_cuda.search(x) for x in possible_cudas]
possible_cudas = [x.group(1) for x in possible_cudas if x is not None]
# Try linking cuda folder, or everything in local
if len(possible_cudas) == 0:
os.system(f"ldconfig /usr/local/")
else:
find_number = re.compile(r"([\d\.]{2,})")
latest_cuda = np.argsort([float(find_number.search(x).group(1)) for x in possible_cudas])[::-1][0]
latest_cuda = possible_cudas[latest_cuda]
os.system(f"ldconfig /usr/local/{latest_cuda}")
pass
importlib.reload(bnb)
importlib.reload(triton)
try:
libcuda_dirs = lambda: None
if Version(triton.__version__) >= Version("3.0.0"):
try: from triton.backends.nvidia.driver import libcuda_dirs
except: pass
else: from triton.common.build import libcuda_dirs
cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
libcuda_dirs()
except:
warnings.warn(
"Unsloth: CUDA is not linked properly.\n"\
"Try running `python -m bitsandbytes` then `python -m xformers.info`\n"\
"We tried running `ldconfig /usr/lib64-nvidia` ourselves, but it didn't work.\n"\
"You need to run in your terminal `sudo ldconfig /usr/lib64-nvidia` yourself, then import Unsloth.\n"\
"Also try `sudo ldconfig /usr/local/cuda-xx.x` - find the latest cuda version.\n"\
"Unsloth will still run for now, but maybe it might crash - let's hope it works!"
)
pass
from .models import *
from .save import *
from .chat_templates import *
from .tokenizer_utils import *
from .trainer import *
build/lib/build/lib/unsloth/chat_templates.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
__all__ = [
"get_chat_template",
"test_chat_templates",
"test_hf_gguf_equivalence",
"remove_special_tokens",
"to_sharegpt",
"standardize_sharegpt",
"apply_chat_template",
"train_on_responses_only",
"test_construct_chat_template",
]
from transformers import StoppingCriteria, StoppingCriteriaList
from torch import LongTensor, FloatTensor
from transformers.models.llama.modeling_llama import logger
from .save import patch_saving_functions
import os
import shutil
from .tokenizer_utils import *
from .models._utils import patch_tokenizer
import re
CHAT_TEMPLATES = {}
# =========================================== Unsloth
# Unsloth efficient template leverages from Zephyr
unsloth_template = \
"{{ bos_token }}"\
"{% if messages[0]['role'] == 'system' %}"\
"{{ messages[0]['content'] + '\n' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% else %}"\
"{{ 'You are a helpful assistant to the user\n' }}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '>>> User: ' + message['content'] + '\n' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ '>>> Assistant: ' + message['content'] + eos_token + '\n' }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '>>> Assistant: ' }}"\
"{% endif %}"
pass
unsloth_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}{{ .System }}
{{ end }}{{ if .Prompt }}>>> User: {{ .Prompt }}
{{ end }}>>> Assistant: {{ .Response }}{__EOS_TOKEN__}
"""
PARAMETER stop "{__EOS_TOKEN__}"
SYSTEM """You are a helpful assistant to the user"""
'''
unsloth_eos_token = "eos_token"
CHAT_TEMPLATES["unsloth"] = (unsloth_template, unsloth_eos_token, False, unsloth_ollama,)
pass
# =========================================== Zephyr
# Zephyr has no BOS!
zephyr_template = \
"{% for message in messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '<|user|>\n' + message['content'] + eos_token + '\n' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ '<|assistant|>\n' + message['content'] + eos_token + '\n' }}"\
"{% else %}"\
"{{ '<|system|>\n' + message['content'] + eos_token + '\n' }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '<|assistant|>\n' }}"\
"{% endif %}"
pass
zephyr_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}<|system|>
{{ .System }}{__EOS_TOKEN__}
{{ end }}{{ if .Prompt }}<|user|>
{{ .Prompt }}{__EOS_TOKEN__}
{{ end }}<|assistant|>
{{ .Response }}{__EOS_TOKEN__}
"""
PARAMETER stop "{__EOS_TOKEN__}"
'''
zephyr_eos_token = "eos_token"
CHAT_TEMPLATES["zephyr"] = (zephyr_template, zephyr_eos_token, False, zephyr_ollama,)
pass
# =========================================== ChatML
# ChatML has no BOS and not EOS! Rather <|im_start|> and <|im_end|> acts as BOS / EOS.
chatml_template = \
"{% for message in messages %}"\
"{% if message['role'] == 'user' %}"\
"{{'<|im_start|>user\n' + message['content'] + '<|im_end|>\n'}}"\
"{% elif message['role'] == 'assistant' %}"\
"{{'<|im_start|>assistant\n' + message['content'] + '<|im_end|>\n' }}"\
"{% else %}"\
"{{ '<|im_start|>system\n' + message['content'] + '<|im_end|>\n' }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '<|im_start|>assistant\n' }}"\
"{% endif %}"
pass
chatml_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>
"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
'''
chatml_eos_token = "<|im_end|>"
CHAT_TEMPLATES["chatml"] = (chatml_template, chatml_eos_token, True, chatml_ollama,)
pass
# =========================================== Mistral-1
# Mistral Instruct doesn't allow system prompts, so we append it to the user message.
mistral_template = \
"{{ bos_token }}"\
"{% if messages[0]['role'] == 'system' %}"\
"{% if messages[1]['role'] == 'user' %}"\
"{{ '[INST] ' + messages[0]['content'] + ' ' + messages[1]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[2:] %}"\
"{% else %}"\
"{{ '[INST] ' + messages[0]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% endif %}"\
"{% else %}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '[INST] ' + message['content'] + ' [/INST]' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ message['content'] + eos_token }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"
pass
# Ollama from https://www.ollama.com/library/mistral
mistral_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """[INST] {{ if .System }}{{ .System }} {{ end }}{{ .Prompt }} [/INST]"""
PARAMETER stop "{__EOS_TOKEN__}"
'''
mistral_eos_token = "eos_token"
CHAT_TEMPLATES["mistral"] = (mistral_template, mistral_eos_token, False, mistral_ollama,)
pass
# =========================================== Llama-2
# Adds BOS to every convo! And weird <<SYS>> system messages.
llama_template = \
"{% if messages[0]['role'] == 'system' %}"\
"{% if messages[1]['role'] == 'user' %}"\
"{{ bos_token + '[INST] <<SYS>>\n' + messages[0]['content'] + '\n<</SYS>>\n\n' + messages[1]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[2:] %}"\
"{% else %}"\
"{{ bos_token + '[INST] ' + messages[0]['content'] + ' [/INST]' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% endif %}"\
"{% else %}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ bos_token + '[INST] ' + message['content'].strip() + ' [/INST]' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ ' ' + message['content'].strip() + ' ' + eos_token }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"
pass
# Ollama from https://www.ollama.com/library/llama3
llama_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """[INST] <<SYS>>{{ .System }}<</SYS>>
{{ .Prompt }} [/INST]"""
PARAMETER stop "{__EOS_TOKEN__}"
'''
llama_eos_token = "eos_token"
CHAT_TEMPLATES["llama"] = (llama_template, llama_eos_token, False, llama_ollama,)
pass
# =========================================== Vicuna
# https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md#prompt-template
vicuna_template = \
"{{ bos_token }}"\
"{% if messages[0]['role'] == 'system' %}"\
"{{ messages[0]['content'] + ' ' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% else %}"\
"{{ 'A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user\\'s questions.' + ' ' }}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ 'USER: ' + message['content'] + ' ' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ 'ASSISTANT: ' + message['content'] + eos_token }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ 'ASSISTANT:' }}"\
"{% endif %}"
pass
# Ollama from https://www.ollama.com/library/vicuna
vicuna_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}{{ .System }} {{ end }}{{ if .Prompt }}USER: {{ .Prompt }} {{ end }}ASSISTANT: {{ .Response }} {__EOS_TOKEN__}"""
PARAMETER stop "{__EOS_TOKEN__}"
'''
vicuna_eos_token = "eos_token"
CHAT_TEMPLATES["vicuna"] = (vicuna_template, vicuna_eos_token, False, vicuna_ollama,)
pass
# =========================================== Vicuna Old
# https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md#prompt-template
vicuna_old_template = \
"{{ bos_token }}"\
"{% if messages[0]['role'] == 'system' %}"\
"{{ messages[0]['content'] + '\n' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% else %}"\
"{{ 'A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human\\'s questions.' + '\n' }}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '### Human: ' + message['content'] + '\n' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ '### Assistant: ' + message['content'] + eos_token + '\n' }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '### Assistant:' }}"\
"{% endif %}"
pass
vicuna_old_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}{{ .System }}
{{ end }}{{ if .Prompt }}### Human: {{ .Prompt }}
{{ end }}### Assistant: {{ .Response }}{__EOS_TOKEN__}
"""
PARAMETER stop "{__EOS_TOKEN__}"
SYSTEM """A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."""
'''
vicuna_old_eos_token = "eos_token"
CHAT_TEMPLATES["vicuna_old"] = (vicuna_old_template, vicuna_old_eos_token, False, vicuna_old_ollama,)
pass
# =========================================== Alpaca multi turn
# https://github.com/tatsu-lab/stanford_alpaca Changed for multi-turn convos
alpaca_template = \
"{{ bos_token }}"\
"{% if messages[0]['role'] == 'system' %}"\
"{{ messages[0]['content'] + '\n\n' }}"\
"{% set loop_messages = messages[1:] %}"\
"{% else %}"\
"{{ 'Below are some instructions that describe some tasks. Write responses that appropriately complete each request.\n\n' }}"\
"{% set loop_messages = messages %}"\
"{% endif %}"\
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '### Instruction:\n' + message['content'] + '\n\n' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ '### Response:\n' + message['content'] + eos_token + '\n\n' }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '### Response:\n' }}"\
"{% endif %}"
pass
alpaca_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}{{ .System }}
{{ end }}{{ if .Prompt }}### Instruction:
{{ .Prompt }}{{ end }}
### Response:
{{ .Response }}{__EOS_TOKEN__}
"""
PARAMETER stop "{__EOS_TOKEN__}"
SYSTEM """Below are some instructions that describe some tasks. Write responses that appropriately complete each request."""
'''
alpaca_eos_token = "eos_token"
CHAT_TEMPLATES["alpaca"] = (alpaca_template, alpaca_eos_token, False, alpaca_ollama,)
pass
# =========================================== Gemma
# https://huggingface.co/google/gemma-7b-it
# Notice we must use |trim for lstrip and rstrip. <start_of_turn> maps to 106.
# <end_of_turn> maps to 107. user and model are normal 1 word tokens.
gemma_template = \
"{{ bos_token }}"\
"{% if messages[0]['role'] == 'system' %}"\
"{{'<start_of_turn>user\n' + messages[0]['content'] | trim + ' ' + messages[1]['content'] | trim + '<end_of_turn>\n'}}"\
"{% set loop_messages = messages[2:] %}"\
"{% endif %}"\
"{% for message in messages %}"\
"{% if message['role'] == 'user' %}"\
"{{'<start_of_turn>user\n' + message['content'] | trim + '<end_of_turn>\n'}}"\
"{% elif message['role'] == 'assistant' %}"\
"{{'<start_of_turn>model\n' + message['content'] | trim + '<end_of_turn>\n' }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '<start_of_turn>model\n' }}"\
"{% endif %}"
pass
# Ollama from https://www.ollama.com/library/gemma
gemma_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """<start_of_turn>user
{{ if .System }}{{ .System }} {{ end }}{{ .Prompt }}<end_of_turn>
<start_of_turn>model
{{ .Response }}<end_of_turn>
"""
PARAMETER repeat_penalty 1
PARAMETER stop "<start_of_turn>"
PARAMETER stop "<end_of_turn>"
PARAMETER penalize_newline false
'''
gemma_eos_token = "<end_of_turn>"
CHAT_TEMPLATES["gemma"] = (gemma_template, gemma_eos_token, True, gemma_ollama,)
pass
# =========================================== Gemma with ChatML instead
# We find using <eos> is still more appropriate!
gemma_chatml_template = "{{ bos_token }}" + chatml_template
pass
gemma_chatml_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>
"""
PARAMETER repeat_penalty 1
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER penalize_newline false
'''
gemma_chatml_eos_token = (
{"<start_of_turn>" : "<|im_start|>", "<eos>" : "<|im_end|>"},
"<|im_end|>",
)
CHAT_TEMPLATES["gemma_chatml"] = (gemma_chatml_template, gemma_chatml_eos_token, True, gemma_chatml_ollama,)
pass
# =========================================== Gemma 2
# Same as Gemma 1, but with sliding window attention!
# https://ollama.com/library/gemma2/blobs/6522ca797f47
gemma2_template = gemma_template
gemma2_ollama = gemma_ollama + "PARAMETER num_ctx 4096\n"
gemma2_eos_token = "<end_of_turn>"
CHAT_TEMPLATES["gemma2"] = (gemma2_template, gemma2_eos_token, True, gemma2_ollama,)
# =========================================== Gemma 2 with ChatML instead
gemma2_chatml_template = gemma_chatml_template
gemma2_chatml_ollama = gemma_chatml_ollama + "PARAMETER num_ctx 4096\n"
gemma2_chatml_eos_token = gemma_chatml_eos_token
CHAT_TEMPLATES["gemma2_chatml"] = (gemma2_chatml_template, gemma2_chatml_eos_token, True, gemma2_chatml_ollama,)
pass
# =========================================== Llama-3
# Weirdly \n\n is needed?
llama3_template = \
"{{ bos_token }}"\
"{% for message in messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ '<|start_header_id|>user<|end_header_id|>\n\n' + message['content'] | trim + '<|eot_id|>' }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' + message['content'] | trim + '<|eot_id|>' }}"\
"{% else %}"\
"{{ '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n' + message['content'] | trim + '<|eot_id|>' }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}"\
"{% endif %}"
pass
# Ollama from https://www.ollama.com/library/llama3
llama3_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
{{ .Response }}<|eot_id|>"""
PARAMETER stop "<|start_header_id|>"
PARAMETER stop "<|end_header_id|>"
PARAMETER stop "<|eot_id|>"
'''
llama3_template_eos_token = "eos_token"
CHAT_TEMPLATES["llama-3"] = (llama3_template, llama3_template_eos_token, False, llama3_ollama,)
pass
# =========================================== Phi-3
phi3_template = \
"{{ bos_token }}"\
"{% for message in messages %}"\
"{% if message['role'] == 'user' %}"\
"{{'<|user|>\n' + message['content'] + '<|end|>\n'}}"\
"{% elif message['role'] == 'assistant' %}"\
"{{'<|assistant|>\n' + message['content'] + '<|end|>\n'}}"\
"{% else %}"\
"{{'<|' + message['role'] + '|>\n' + message['content'] + '<|end|>\n'}}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '<|assistant|>\n' }}"\
"{% endif %}"
pass
# Ollama from https://www.ollama.com/library/phi3
phi3_ollama = \
'''
FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}<|system|>
{{ .System }}<|end|>
{{ end }}{{ if .Prompt }}<|user|>
{{ .Prompt }}<|end|>
{{ end }}<|assistant|>
{{ .Response }}<|end|>
"""
PARAMETER stop "<|end|>"
PARAMETER stop "<|user|>"
PARAMETER stop "<|assistant|>"
'''
phi3_template_eos_token = "<|end|>"
CHAT_TEMPLATES["phi-3"] = (phi3_template, phi3_template_eos_token, False, phi3_ollama,)
pass
def get_chat_template(
tokenizer,
chat_template = "chatml",
mapping = {"role" : "role", "content" : "content", "user" : "user", "assistant" : "assistant"},
map_eos_token = True,
system_message = None,
):
assert(type(map_eos_token) is bool)
old_tokenizer = tokenizer
IS_GEMMA = False
if tokenizer.__class__.__name__.startswith("Gemma"):
if chat_template == "chatml": chat_template = "gemma_chatml"
IS_GEMMA = True
pass
# We add a check for Llama-3
# if chat_template == "llama-3":
# tokenizer._using_llama3_template = True
# else:
# llama3_tokens = set(["<|end_header_id|>", "<|eot_id|>", "<|start_header_id|>"])
# check_llama3_tokens = llama3_tokens & set(str(x) for x in tokenizer.added_tokens_decoder.values())
# if len(check_llama3_tokens) == len(llama3_tokens):
# tokenizer._using_llama3_template = True
# pass
# pass
# We first check if the tokenizer is a fast one. If not, we cannot convert this!
is_fast_tokenizer = getattr(tokenizer, "is_fast", False)
old_padding_side = tokenizer.padding_side
same_padding_token = False
if type(chat_template) in (list, tuple,):
chat_template, stop_word = chat_template
assert(type(chat_template) is str)
assert(type(stop_word) is str)
ollama_modelfile = None
elif type(chat_template) is str:
chat_template, stop_word, yes_map_eos_token, ollama_modelfile = CHAT_TEMPLATES[chat_template]
# Check mapping to eos_token
if not map_eos_token and yes_map_eos_token: map_eos_token = True
if not yes_map_eos_token and map_eos_token: map_eos_token = False
if type(stop_word) in (list, tuple,):
token_mapping, stop_word = stop_word
assert(type(token_mapping) is dict)
else:
token_mapping = None
assert(type(stop_word) is str)
# Check fast tokenizer
if not is_fast_tokenizer:
print(
f"Unsloth: Not a fast tokenizer, so can't process it as of yet :(\n"\
"Please log a Github issue if you want this as a new feature!\n"\
"Your chat template will still work, but it won't add or edit tokens."
)
elif token_mapping is not None:
# token_mapping = {"<start_of_turn>" : "<|im_start|>", "<end_of_turn>" : "<|im_end|>"}
# For Gemma :)
string_vocab = tokenizer._tokenizer.to_str()
skipped = 0
for old_token, new_token in token_mapping.items():
old_count = string_vocab.count(f'"{old_token}"')
new_count = string_vocab.count(f'"{new_token}"')
if new_count != 0:
print(f"{new_token} is already a token. Skipping.")
skipped += 1
elif old_count == 0:
raise RuntimeError(f"{old_token} was not part of the tokenizer!")
else:
string_vocab = string_vocab.replace(f'"{old_token}"', f'"{new_token}"')
pass
pass
if map_eos_token and (not stop_word in token_mapping.values()):
# Do not map 107 = <|im_end|> and 1 = <|im_end|>. This will reduce the vocab size by 1
logger.warning_once(f"Unsloth: Will map {stop_word} to EOS = {tokenizer.eos_token}.")
string_vocab = string_vocab.replace(tokenizer.eos_token, stop_word)
pass
if skipped != len(token_mapping):
new_tokenizer = tokenizer._tokenizer.from_str(string_vocab)
# Careful on pad_token
old_pad_token = tokenizer.pad_token
if old_pad_token == tokenizer.eos_token:
old_pad_token = stop_word
same_padding_token = True
pass
if map_eos_token:
new_tokenizer = tokenizer.__class__(
tokenizer_object = new_tokenizer,
eos_token = stop_word,
pad_token = old_pad_token,
)
else:
new_tokenizer = tokenizer.__class__(
tokenizer_object = new_tokenizer,
pad_token = old_pad_token,
)
pass
# Must fix the sentence piece tokenizer since there's no tokenizer.model file!
tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,)
else:
pass
elif map_eos_token and (stop_word != "eos_token"):
logger.warning_once(f"Unsloth: Will map {stop_word} to EOS = {tokenizer.eos_token}.")
# Replaces the old EOS token with a new one.
# Useful for ChatML <|im_end|> for example.
# Usually we train 2 more tokens <|im_start|> and <|im_end|>
# But training the lm_head and embeddings are slow!
# This is a HACK!
# Idea from https://huggingface.co/cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser
old_bos_token = getattr(tokenizer, "bos_token", None)
old_eos_token = getattr(tokenizer, "eos_token", None)
old_pad_token = getattr(tokenizer, "pad_token", None)
old_unk_token = getattr(tokenizer, "unk_token", None)
string_vocab = tokenizer._tokenizer.to_str()
# First check if new stop_word is in the tokenizer
if stop_word in string_vocab:
# We shall swap them around
temporary_stop_token = "<|:__TEMP//STOP//TOKEN__:|>"
string_vocab = string_vocab.replace(old_eos_token, temporary_stop_token)
string_vocab = string_vocab.replace(stop_word, old_eos_token)
string_vocab = string_vocab.replace(temporary_stop_token, stop_word)
else:
string_vocab = string_vocab.replace(old_eos_token, stop_word)
pass
new_tokenizer = tokenizer._tokenizer.from_str(string_vocab)
# Careful on pad_token
if old_pad_token == old_eos_token:
old_pad_token = stop_word
same_padding_token = True
pass
new_tokenizer = tokenizer.__class__(
tokenizer_object = new_tokenizer,
bos_token = old_bos_token,
eos_token = stop_word,
unk_token = old_unk_token,
pad_token = old_pad_token,
)
# Must fix the sentence piece tokenizer since there's no tokenizer.model file!
token_mapping = { old_eos_token : stop_word, }
tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,)
pass
else:
raise TypeError(
f"Unsloth: `chat_template` must be a tuple of (your_template, eos_token,) or one of\n"\
f"{CHAT_TEMPLATES.keys()}"
)
pass
# For ShareGPT role -> from and content -> value
chat_template = chat_template\
.replace("'role'", "'" + mapping["role"] + "'")\
.replace("'content'", "'" + mapping["content"] + "'")\
.replace("'user'", "'" + mapping["user"] + "'")\
.replace("'assistant'", "'" + mapping["assistant"] + "'")
# Careful on Gemma
# bos_token is a must or else losses become too high
if IS_GEMMA and not chat_template.startswith("{{ bos_token }}"):
chat_template = "{{ bos_token }}" + chat_template
pass
_, tokenizer = patch_tokenizer(model = None, tokenizer = tokenizer)
tokenizer.padding_side = old_padding_side
tokenizer.chat_template = chat_template
# Also fix up other tokens
old_pad_token = getattr(old_tokenizer, "pad_token", None)
old_bos_token = getattr(old_tokenizer, "bos_token", None)
old_unk_token = getattr(old_tokenizer, "unk_token", None)
new_pad_token = getattr(tokenizer, "pad_token", None)
new_bos_token = getattr(tokenizer, "bos_token", None)
new_unk_token = getattr(tokenizer, "unk_token", None)
if old_bos_token != new_bos_token: tokenizer.bos_token = old_bos_token
if old_unk_token != new_unk_token: tokenizer.unk_token = old_unk_token
if not same_padding_token:
if old_pad_token != new_pad_token: tokenizer.pad_token = old_pad_token
pass
# stopping_criteria = create_stopping_criteria(tokenizer, stop_word)
# Patch saving functions
tokenizer = patch_saving_functions(tokenizer)
# Add Ollama
tokenizer._ollama_modelfile = ollama_modelfile
tokenizer._system_message = system_message
return tokenizer#, stopping_criteria
pass
def remove_special_tokens(tokenizer, prompt):
# Removes double BOS token
if prompt.startswith(tokenizer.bos_token):
prompt = prompt[len(tokenizer.bos_token):]
pass
return prompt
pass
def _parse_combined_prompt(combined_prompt, dataset):
# Find {...}
possible_columns = re.findall(r"\{(.+?)\}", combined_prompt)
dataset_columns = set(dataset.column_names)
for column in possible_columns:
if column not in dataset_columns:
raise KeyError(
f"Unsloth: Your prompt includes '{column}' but this does not exist in the dataset. "\
f"Only allowed columns are {list(dataset_columns)}"
)
pass
pass
# Find [[...]]
optional_prompts = list(re.finditer(r"\[\[.+?\]\]", combined_prompt, flags = re.DOTALL | re.MULTILINE))
optional_prompts = [(x.span(), x.group(0)) for x in optional_prompts]
final_optional_prompts = []
if len(optional_prompts) != 0:
# Add left
left = optional_prompts[0]
l = left[0][0]
if l != 0: final_optional_prompts.append(combined_prompt[:l])
# Add in between
for left, right in zip(optional_prompts[:-1], optional_prompts[1:]):
l, r = left[0][-1], right[0][0]
final_optional_prompts.append(left)
if l != r: final_optional_prompts.append(combined_prompt[l : r])
pass
final_optional_prompts.append(optional_prompts[-1])
# Add right
right = optional_prompts[-1]
r = right[0][1]
if r != len(combined_prompt): final_optional_prompts.append(combined_prompt[r:])
else:
# Just add in the entire string
final_optional_prompts.append(combined_prompt)
pass
check_combined = "".join(x if type(x) is str else x[1] for x in final_optional_prompts)
assert(combined_prompt == check_combined)
return possible_columns, final_optional_prompts
pass
def _create_formatter(possible_columns, final_optional_prompts, user_column_name):
# Start final prompt!
function = ["def __combined_prompt_processor__(examples):"]
columns = list(set(possible_columns))
for column in columns:
function.append(f"{' '*4}{column}__ = examples['{column}']")
function.append(f"{' '*4}texts = []")
function.append(f"{' '*4}for ({', '.join(columns)}) in zip({', '.join(f'{x}__' for x in columns)}):")
# Add optional tags as well!
final_prompt = ""
formatter = []
for j, optional_prompt in enumerate(final_optional_prompts):
if type(optional_prompt) is str:
columns = re.findall(r"\{(.+?)\}", optional_prompt)
formatter += columns
# Must escape \n \r
final_prompt += optional_prompt.encode("unicode-escape").decode("utf-8").replace("'", "\\'").replace('"', '\\"')
else:
where, prompt = optional_prompt
# Strip [[...]]
# Must escape \n \r
prompt = prompt[2:-2].encode("unicode-escape").decode("utf-8").replace("'", "\\'").replace('"', '\\"')
columns = re.findall(r"\{(.+?)\}", prompt)
x = f"__optional_{j}__"
prompt = f"{' '*8}{x} = '{prompt}'.format({', '.join(f'{x} = {x}' for x in columns)}) if {columns[0]} else ''"
function.append(prompt)
formatter.append(x)
final_prompt += "{" + x + "}"
pass
pass
function.insert(1, f"{' '*4}__combined_prompt__ = '{final_prompt}'")
function.append(f"{' '*8}texts.append("\
f"__combined_prompt__.format({', '.join(f'{x} = {x}' for x in formatter)}))")
function.append(f"{' '*4}return " + "{ " + f"'{user_column_name}' : texts" + " }")
return "\n".join(function)
pass
def to_sharegpt(
dataset,
merged_prompt = "",
merged_column_name = "instruction",
output_column_name = "output",
remove_unsued_columns = True,
conversation_extension = 1,
random_state = 3407,
):
"""
Converts a dataset to ShareGPT style.
ShareGPT requires only 1 input and 1 output field.
This means one has to merge multiple columns into 1 for 1 input field.
Use `conversation_extension` to increase the length of each conversation by randomnly
selecting a few and packing them into 1.
merged_prompt = "", Prompt to merge columns into 1 input
merged_column_name = "instruction", Final column name for the input field
output_column_name = "output", Final column name for the output field
remove_unsued_columns = True,
conversation_extension = 1, Automatically combines `conversation_extension` convos into 1
random_state = 3407,
"""
if "conversations" in dataset.column_names:
convo = dataset[0]["conversations"]
if type(convo) is list:
raise TypeError("Unsloth: Your dataset is probably already in ShareGPT format!")
pass
pass
possible_columns, final_optional_prompts = _parse_combined_prompt(merged_prompt, dataset)
function = _create_formatter(possible_columns, final_optional_prompts, merged_column_name)
exec(function, globals())
dataset = dataset.map(__combined_prompt_processor__, batched = True, desc = "Merging columns")
def __convert_to_sharegpt__(examples):
users = examples[merged_column_name]
assistants = examples[output_column_name]
texts = [
[
{"from" : "user", "content" : str(user) },
{"from" : "assistant", "content" : str(assistant)},
] \
for user, assistant in zip(users, assistants)
]
return { "conversations" : texts, }
pass
dataset = dataset.map(
__convert_to_sharegpt__,
batched = True,
desc = "Converting to ShareGPT",
# Remove unsued columns!
remove_columns = dataset.column_names if remove_unsued_columns else None,
)
# Randomnly concat conversations to create a long stream!
from datasets import concatenate_datasets
n_extensions = max(conversation_extension-1, 0)
if n_extensions == 0: return dataset
dataset = dataset.rename_columns({"conversations" : f"conversations0"})
all_shuffled = [dataset]
for j in range(1, n_extensions+1):
shuffled = dataset.shuffle(seed = random_state+j).rename_columns({"conversations0" : f"conversations{j}"})
all_shuffled.append(shuffled)
pass
dataset = concatenate_datasets(all_shuffled, axis = 1)
# Combine them into 1
function = "def __combine_conversations__(examples):\n"
n_extensions += 1
for j in range(n_extensions):
function += f"{' '*4}conversations{j}__ = examples['conversations{j}']\n"
function += f"{' '*4}convos = []\n"
function += f"{' '*4}for ({', '.join(f'conversations{j}' for j in range(n_extensions))}) "\
f"in zip({', '.join(f'conversations{j}__' for j in range(n_extensions))}):\n"
function += f"{' '*8}convos.append("\
f"{'+'.join(f'conversations{j}' for j in range(n_extensions))})\n"
function += f"{' '*4}return " + "{ " + f"'conversations' : convos" + " }"
# Map function
exec(function, globals())
dataset = dataset.map(
__combine_conversations__,
batched = True,
desc = "Extending conversations",
# Remove unsued columns!
remove_columns = dataset.column_names if remove_unsued_columns else None,
)
return dataset
pass
def standardize_sharegpt(
dataset,
aliases_for_system = ["system",],
aliases_for_user = ["user", "human", "input",],
aliases_for_assistant = ["gpt", "assistant", "output",],
):
"""
Standardizes ShareGPT and other formats to user/assistant Hugging Face format.
Get aliases for the system, user and assistant roles.
These shall map to "system", "user" and "assistant" respectively.
aliases_for_system = ["system",],
aliases_for_user = ["user", "human", "input",],
aliases_for_assistant = ["gpt", "assistant", "output",],
"""
import collections
import itertools
convos = dataset[:10]["conversations"]
uniques = collections.defaultdict(list)
for convo in convos:
for message in convo:
for key, value in message.items():
uniques[key].append(value)
pass
# Must be only 2 entries
assert(len(uniques.keys()) == 2)
keys = list(uniques.keys())
length_first = len(set(uniques[keys[0]]))
length_second = len(set(uniques[keys[1]]))
if length_first < length_second:
# Role is assigned to the first element
role_key = keys[0]
content_key = keys[1]
else:
role_key = keys[1]
content_key = keys[0]
pass
# Check roles are in aliases
all_aliases = set(aliases_for_system + aliases_for_user + aliases_for_assistant)
roles = set(uniques[role_key])
leftover_aliases = (all_aliases | roles) - all_aliases
if len(leftover_aliases) != 0:
raise TypeError(
f"Unsloth: {list(leftover_aliases)} are not in aliases. Please update aliases."
)
pass
# Mapping for aliases
aliases_mapping = {}
for x in aliases_for_system: aliases_mapping[x] = "system"
for x in aliases_for_user: aliases_mapping[x] = "user"
for x in aliases_for_assistant: aliases_mapping[x] = "assistant"
def _standardize_dataset(examples):
convos = examples["conversations"]
all_convos = []
for convo in convos:
new_convo = [
{ "role" : aliases_mapping[message[role_key]], "content" : message[content_key], }
for message in convo
]
all_convos.append(new_convo)
pass
return { "conversations" : all_convos, }
pass
return dataset.map(_standardize_dataset, batched = True, desc = "Standardizing format")
pass
def get_ollama_eos_tokens(tokenizer, extra_eos_tokens = []):
added_tokens_decoder = tokenizer.added_tokens_decoder.values()
added_tokens_decoder = [str(x) for x in added_tokens_decoder]
# Remove added_tokens_decoder duplicates
added_tokens_decoder = list(set(added_tokens_decoder) - set(extra_eos_tokens))
# Remove BOS
if getattr(tokenizer, "bos_token", None) is not None:
added_tokens_decoder = [x for x in added_tokens_decoder if x != tokenizer.bos_token]
pass
repeatted_tokens = []
# Join all vocab
joined_text = "\x01\x00".join(added_tokens_decoder)
for token in added_tokens_decoder:
n = len(token)
repeatted_counts = joined_text.count(token[:n//2])
# Try finding longer than 1/2 of the token in the rest
# For eg <|reserved_special_token_0|>, <|reserved_special_token_1|>
if repeatted_counts > 2:
for j in range(n//2+1, n):
if joined_text.count(token[:j]) < repeatted_counts:
j -= 1
# Remove repeatted tokens to reduce search space
joined_text = joined_text.replace(token[:j], "")
repeatted_tokens.append(token[:j])
break
pass
pass
pass
# Remove duplicates
splitted = joined_text.split("\x01\x00")
final_eos_tokens = []
for old, new in zip(added_tokens_decoder, splitted):
if old == new: final_eos_tokens.append(old)
pass
final_eos_tokens += extra_eos_tokens
final_eos_tokens += repeatted_tokens
# Remove new lines, spaces and HTML tags
filtered_eos_tokens = []
for token in final_eos_tokens:
if token.count("\n") == len(token): continue
elif token.count("▁") == len(token): continue
elif token.startswith("<") and len(token) <= 2: continue
elif token.startswith("</") and len(token) == 3: continue
filtered_eos_tokens.append(token)
pass
return filtered_eos_tokens
pass
def construct_chat_template( \
tokenizer = None,
chat_template = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{SYSTEM}<|eot_id|><|start_header_id|>user<|end_header_id|>
{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{OUTPUT}<|eot_id|><|start_header_id|>user<|end_header_id|>
{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{OUTPUT}<|eot_id|>""",
default_system_message = \
"Below are some instructions that describe some tasks. Write responses that appropriately complete each request.",
extra_eos_tokens = None,
):
"""
Creates a Ollama modelfile and a HF Jinja template from a custom
template. You must provide 2x examples of an input & output.
There is an optional system message as well.
You must use {INPUT}, {OUTPUT} twice, and {SYSTEM} is optional.
"""
# Strip only the left
chat_template = chat_template.lstrip()
assert(tokenizer is not None)
if extra_eos_tokens is None: extra_eos_tokens = []
elif type(extra_eos_tokens) is str: extra_eos_tokens = [extra_eos_tokens,]
vocab = tokenizer.get_vocab()
for extra_eos in extra_eos_tokens:
assert(type(extra_eos) is str)
if extra_eos not in vocab:
raise ValueError(f"Unsloth: `{extra_eos}` is not a singular token in the tokenizer.")
pass
pass
error_msg = \
"Unsloth: Your prompt template must have 2 examples showing the user input {INPUT} "\
"and the assistant output {OUTPUT}\n\n"\
"For example what is not allowed is just:\n"\
"### Input:\\n{INPUT}\\n\\n### Response:\\n{OUTPUT}\\n\n\n"\
"What is required is 2x of this:\n"\
"### Input:\\n{INPUT}\\n\\n### Response:\\n{OUTPUT}\\n"\
"### Input:\\n{INPUT}\\n\\n### Response:\\n{OUTPUT}\\n"
# Check for EOS after {OUTPUT}
if tokenizer.eos_token is not None:
extra_eos_tokens.insert(0, tokenizer.eos_token)
if len(extra_eos_tokens) == 0:
raise RuntimeError(
"Unsloth: Your tokenizer does not have an EOS token? Please provide one via extra_eos_tokens!"
)
pass
# Check tokenizer types
tokenizer_name = tokenizer.name_or_path.lower()
if tokenizer_name.startswith(("unsloth/llama-3-8b-instruct", "unsloth/llama-3-70b-instruct")):
# Add <|eot_id|>
extra_eos_tokens.append("<|eot_id|>")
elif ("<|eot_id|>" in extra_eos_tokens or "<|eot_id|>" in chat_template) and \
tokenizer_name.startswith(("unsloth/llama-3-8b", "unsloth/llama-3-70b")):
# Warn
logger.warning(
"Unsloth: Base llama-3 models did not train <|eot_id|>.\n"\
"Please use the instruct version or use <|end_of_text|>"
)
pass
extra_eos_tokens = list(set(extra_eos_tokens))
count_eos = 0
for eos in extra_eos_tokens:
count_eos += len(re.findall(r"{OUTPUT}" + re.escape(eos), chat_template))
pass
# This forces you to provide 2 input and outputs
final_combined_check = False
try:
# O(N^2) search finding 2 repeatted pieces of text
j = len(chat_template)-1
at_least_one = False
while j > 0:
found = chat_template.rfind(chat_template[j:], 0, j)
if found == -1: break
j -= 1
at_least_one = True
pass
if j > 0: j += 1
else: raise RuntimeError(error_msg)
if not at_least_one: raise RuntimeError(error_msg)
# Must be equivalent to left
final_combined_check = True
# Repeatted text
instruction_response = chat_template[j:]
if instruction_response.count("{INPUT}") != 1 or instruction_response.count("{OUTPUT}") != 1:
raise RuntimeError(error_msg)
pass
# 1st System, Instruction, Output pair
left = chat_template[:j]
# 2nd Instruction, Output pair
right = chat_template[j:]
final_combined_check = left if final_combined_check else chat_template
# Isolate input
extra_eos_tokens_regex = "|".join(f"(?:{re.escape(x)})" for x in extra_eos_tokens)
if len(extra_eos_tokens_regex) != 0:
find_end = f"(?:{extra_eos_tokens_regex})?"
else:
find_end = ""
find_end = r"\{INPUT\}[\s\n]{0,}" + find_end
input_end = list(re.finditer(find_end, right))
assert(len(input_end) == 1)
input_end = input_end[0]
input_end = input_end.span(0)[1]
input_part = right[:input_end]
# Isolate output
output_part = right[input_end:]
# Isolate system
where_system = left.find(input_part)
system_part = left[:where_system if where_system != -1 else len(left)]
# Check if the user provided a correct prompt
combined = system_part + input_part + output_part
if combined != final_combined_check:
combined_changed = combined .replace('\n', '\\n')
left_changed = final_combined_check.replace('\n', '\\n')
raise RuntimeError(
"Unsloth: The prompt template you provided isn't correct. You gave:\n"\
f"{combined_changed}\n\n"\
"But we require the following:\n"\
f"{left_changed}"
)
pass
except:
ending = chat_template[chat_template.find("{OUTPUT}") + len("{OUTPUT}"):]
ending = re.escape(ending)
find_text = "{INPUT}" + ending + "(.+?{OUTPUT}" + ending + ")"
response_part = re.findall(find_text, chat_template, flags = re.DOTALL | re.MULTILINE)
response_part = response_part[0]
for j in range(1, len(response_part)):
try_find = re.escape(response_part[:j])
try: found = next(re.finditer("(" + try_find + ").+?\{INPUT\}", chat_template, flags = re.DOTALL | re.MULTILINE))
except: break
pass
separator = found.group(1)
response_start = chat_template.find(response_part)
start_instruction = chat_template[:response_start].rfind(separator)
if start_instruction == -1: start_instruction = 0
instruction_part = chat_template[start_instruction:response_start]
combined = instruction_part + response_part
where = chat_template.find(combined)
system_part = chat_template[:where]
system_part, input_part, output_part = system_part, instruction_part, response_part
pass
if count_eos == 0:
logger.warning("Unsloth: We automatically added an EOS token to stop endless generations.")
eos = extra_eos_tokens[0]
output_part = output_part + eos
pass
# Ollama modelfile parts
# Check bos_token is in system prompt
ollama_system = system_part
has_bos_token = False
always_bos_token = False
if tokenizer("A").input_ids[0] == getattr(tokenizer, "bos_token_id", None):
always_bos_token = True
if ollama_system.startswith(tokenizer.bos_token):
has_bos_token = True
ollama_system = ollama_system[len(tokenizer.bos_token):]
pass
pass
# Check system
if "{SYSTEM}" in ollama_system:
system_modelfile = "{{ if .System }}" + ollama_system.replace("{SYSTEM}", "{{ .System }}") + "{{ end }}"
else:
system_modelfile = ollama_system
pass
input_modelfile = "{{ if .Prompt }}" + input_part .replace("{INPUT}", "{{ .Prompt }}") + "{{ end }}"
output_modelfile = output_part.replace("{OUTPUT}", "{{ .Response }}")
# Ollama EOS
ollama_eos = get_ollama_eos_tokens(tokenizer, extra_eos_tokens)
ollama_eos = '\n'.join(f'PARAMETER stop "{eos}"' for eos in ollama_eos)
# Ollama modelfile
modelfile = 'FROM {__FILE_LOCATION__}\n\n'\
'TEMPLATE """' + system_modelfile + input_modelfile + output_modelfile + \
'"""\n\n' + ollama_eos
# HF Jinja Chat template
def process(part, which, content = "message['content']"):
if part.endswith(which):
part = "'" + part[:part.find(which)] + f"' + {content}"
elif part.startswith(which):
part = f"{content} + '" + part[part.find(which):] + "'"
else:
part = "'" + part.replace(which, f"' + {content} + '") + "'"
if part.startswith("'' + "): part = part[5:]
return part
pass
input_jinja = process(input_part, "{INPUT}")
output_jinja = process(output_part, "{OUTPUT}")
pass
jinja_template = \
"{% for message in loop_messages %}"\
"{% if message['role'] == 'user' %}"\
"{{ " + input_jinja + " }}"\
"{% elif message['role'] == 'assistant' %}"\
"{{ " + output_jinja + " }}"\
"{% else %}"\
"{{ raise_exception('Only user and assistant roles are supported!') }}"\
"{% endif %}"\
"{% endfor %}"\
"{% if add_generation_prompt %}"\
"{{ '" + output_part[:output_part.find("{OUTPUT}")] + "' }}"\
"{% endif %}"
pass
# Now add system prompt to jinja
if len(system_part) != 0:
partial_system = process(system_part, "{SYSTEM}", "messages[0]['content']")
partial_system = partial_system.replace("{SYSTEM}", "")
if "{SYSTEM}" in partial_system:
if default_system_message is None:
raise RuntimeError("Unsloth: Please specify a default system message!")
pass
# Separate the BOS
if has_bos_token:
partial_system = partial_system.replace(tokenizer.bos_token, "", 1)
system_part = system_part .replace(tokenizer.bos_token, "", 1)
pass
partial_system = \
"{% if messages[0]['role'] == 'system' %}"\
"{{ " + partial_system + " }}"\
"{% set loop_messages = messages[1:] %}"
if default_system_message is not None:
full_system = system_part.replace("{SYSTEM}", default_system_message)
if "{SYSTEM}" in system_part:
modelfile += '\nSYSTEM "' + default_system_message + '"'
pass
partial_system += "{% else %}"\
"{{ '" + full_system + "' }}"\
"{% set loop_messages = messages %}"\
"{% endif %}"
else:
partial_system += "{% endif %}"
pass
jinja_template = partial_system + jinja_template
if has_bos_token:
jinja_template = "{{ bos_token }}" + jinja_template
pass
# Fix missing loop_messages
if "{% set loop_messages = messages %}" not in jinja_template:
jinja_template = jinja_template.replace(
"{% for message in loop_messages %}",
"{% for message in messages %}",
1, # Only replace the first one
)
pass
# Check if system part is the same!
jinja_template = re.sub(
r"\{\% if messages\[0\]\['role'\] \=\= 'system' \%\}\{\{ '(.+?)' \}\}"\
r"\{\% set loop\_messages \= messages\[1\:\] \%\}"\
r"\{\% else \%\}\{\{ '\1' \}\}\{\% set loop\_messages \= messages \%\}\{\% endif \%\}"\
r"\{\% for message in loop\_messages \%\}",
r"{{ '\1' }}{% for message in messages %}",
jinja_template, flags = re.MULTILINE | re.DOTALL,
)
# Check jinja tempate for bos
if always_bos_token:
if not jinja_template.startswith("{{ bos_token }}"):
jinja_template = "{{ bos_token }}" + jinja_template
pass
# Get instruction and output parts for train_on_inputs = False
input_part = input_part [:input_part .find("{INPUT}")]
output_part = output_part[:output_part.find("{OUTPUT}")]
return modelfile, jinja_template, input_part, output_part
pass
def test_construct_chat_template():
token = "hf_"
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct", token = token)
chat_template = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{SYSTEM}<|eot_id|><|start_header_id|>user<|end_header_id|>
{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{OUTPUT}<|eot_id|><|start_header_id|>user<|end_header_id|>
{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{OUTPUT}<|eot_id|>"""
default_system_message = \
"Below are some instructions that describe some tasks. Write responses that appropriately complete each request."
extra_eos_tokens = None
modelfile, jinja_template, _, _ = construct_chat_template(
tokenizer = tokenizer,
chat_template = chat_template,
extra_eos_tokens = extra_eos_tokens,
)
messages = [
{"role": "system", "content": "You are an assistant"},
{"role": "user", "content": "What is 2+2?"},
{"role": "assistant", "content": "It's 4."},
{"role": "user", "content": "Ok!"},
{"role": "assistant", "content": "Anything else?"},
{"role": "user", "content": "What's 2x2?"},
]
correct_output = tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
tokenizer.chat_template = jinja_template
new_output = tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
assert(correct_output == new_output)
pass
pass
def apply_chat_template( \
dataset,
tokenizer = None,
chat_template = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{SYSTEM}<|eot_id|><|start_header_id|>user<|end_header_id|>
{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{OUTPUT}<|eot_id|><|start_header_id|>user<|end_header_id|>
{INPUT}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{OUTPUT}<|eot_id|>""",
default_system_message = \
"Below are some instructions that describe some tasks. Write responses that appropriately complete each request.",
extra_eos_tokens = None,
):
"""
Creates a Ollama modelfile and a HF Jinja template from a custom
template. You must provide 2x examples of an input & output.
There is an optional system message as well.
You must use {INPUT}, {OUTPUT} twice, and {SYSTEM} is optional.
"""
modelfile, jinja_template, input_part, output_part = construct_chat_template(
tokenizer = tokenizer,
chat_template = chat_template,
default_system_message = default_system_message,
extra_eos_tokens = extra_eos_tokens,
)
def formatting_prompts_func(examples):
convos = examples["conversations"]
texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
return { "text" : texts, }
pass
tokenizer.chat_template = jinja_template
tokenizer._ollama_modelfile = modelfile
tokenizer._unsloth_input_part = input_part
tokenizer._unsloth_output_part = output_part
return dataset.map(formatting_prompts_func, batched = True,)
pass
def train_on_responses_only(
trainer,
instruction_part = None,
response_part = None,
):
"""
Trains only on responses and not on the instruction by masking out
the labels with -100 for the instruction part.
"""
tokenizer = trainer.tokenizer
if not hasattr(tokenizer, "_unsloth_input_part") or \
not hasattr(tokenizer, "_unsloth_output_part"):
if instruction_part is None or response_part is None:
raise ValueError("Unsloth: instruction_part and response_part must be given!")
pass
elif (instruction_part is not None or response_part is not None) and \
(hasattr(tokenizer, "_unsloth_input_part") or hasattr(tokenizer, "_unsloth_output_part")):
raise ValueError("Unsloth: Your tokenizer already has instruction and response parts set - do not give custom ones!")
else:
instruction_part = tokenizer._unsloth_input_part
response_part = tokenizer._unsloth_output_part
pass
instruction_ids = tokenizer(instruction_part, add_special_tokens = False).input_ids
response_ids = tokenizer(response_part, add_special_tokens = False).input_ids
instruction_length = len(instruction_ids)
response_length = len(response_ids)
max_length = max(instruction_length, response_length)
def _train_on_responses_only(examples):
input_ids_ = examples["input_ids"]
all_labels = []
for input_ids in input_ids_:
labels = [-100] * len(input_ids)
m = len(input_ids) - max_length
first_response = response_ids[0]
first_instruction = instruction_ids[0]
j = 0
while j < m:
if input_ids[j] == first_response:
if input_ids[j : j+response_length] == response_ids:
j = j + response_length
start = j
while j < m:
if input_ids[j] == first_instruction and input_ids[j : j+instruction_length] == instruction_ids:
j = j + instruction_length
labels[start : j] = input_ids[start : j]
break
elif j == (m-1):
j = m
labels[start:] = input_ids[start:]
break
pass
j += 1
pass
pass
pass
j += 1
pass
all_labels.append(labels)
pass
return { "labels" : all_labels }
pass
trainer.train_dataset = trainer.train_dataset.map(_train_on_responses_only, batched = True)
return trainer
pass
def create_stopping_criteria(tokenizer, stop_word = "eos_token"):
class StoppingCriteriaSub(StoppingCriteria):
__slots__ = "stop_token", "single_match", "length",
def __init__(self, stops = "eos_token", device = "cuda", encounters = 1):
super().__init__()
if stops == "eos_token":
self.stop_token = torch.tensor(tokenizer.eos_token_id, device = "cuda")
self.length = 1
else:
self.stop_token = tokenizer(["\n" + stops], add_special_tokens = False, return_tensors = "pt")
self.stop_token = self.stop_token.input_ids.ravel()[1:].to("cuda")
self.length = self.stop_token.shape[0]
pass
self.single_match = self.length == 1
pass
def __call__(self, input_ids: LongTensor, scores: FloatTensor) -> bool:
input_ids = input_ids.ravel()
last_token = input_ids[-1]
if self.single_match and (last_token == self.stop_token): return True
if input_ids.shape[0] >= self.length and \
(input_ids[-self.length:] == self.stop_token).all(): return True
return False
pass
pass
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops = stop_word)])
return stopping_criteria
pass
def test_chat_templates():
messages = [
{"role": "system","content": " You are a friendly chatbot.",},
{"role": "user", "content": "What is 2+2?"},
{"role": "assistant", "content": "It's 4."},
{"role": "user", "content": " But 2+2 is equal to 5. "},
{"role": "assistant", "content": "No I'm sure its 4."},
{"role": "user", "content": " No it's 100% 5! "},
]
# Zephyr
from transformers import AutoTokenizer
template = zephyr_template
correct_tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
correct_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
# Chatml
template = chatml_template
correct_tokenizer = AutoTokenizer.from_pretrained("teknium/OpenHermes-2.5-Mistral-7B")
correct_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
# Mistral
template = mistral_template
correct_tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
correct_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
# Llama
template = llama_template
correct_tokenizer = AutoTokenizer.from_pretrained("unsloth/llama-2-7b-chat")
correct_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
# Vicuna
try:
from fastchat.conversation import get_conv_template
except:
os.system("pip -qqq install git+https://github.com/lm-sys/FastChat.git")
from fastchat.conversation import get_conv_template
correct_prompt = get_conv_template("vicuna_v1.1")
for j in range(len(messages)-1):
correct_prompt.append_message(correct_prompt.roles[j%2==1], messages[j+1]["content"])
correct_prompt.append_message(correct_prompt.roles[1], "")
correct_prompt = tokenizer.bos_token + correct_prompt.get_prompt()
template = vicuna_template
correct_tokenizer = AutoTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
try:
from fastchat.conversation import get_conv_template
except:
os.system("pip -qqq install git+https://github.com/lm-sys/FastChat.git")
from fastchat.conversation import get_conv_template
correct_prompt = get_conv_template("zero_shot")
for j in range(len(messages)-1):
correct_prompt.append_message(correct_prompt.roles[j%2==1], messages[j+1]["content"])
correct_prompt.append_message(correct_prompt.roles[1], "")
correct_prompt = tokenizer.bos_token + correct_prompt.get_prompt()
template = vicuna_old_template
correct_tokenizer = AutoTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
# We add </s> ourselves
assert(correct_prompt == our_prompt.replace("</s>", ""))
# Gemma
correct_tokenizer = AutoTokenizer.from_pretrained("unsloth/gemma-7b-it")
correct_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = gemma_template
our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
assert(our_prompt == correct_prompt)
# Llama-3
template = llama3_template
correct_tokenizer = AutoTokenizer.from_pretrained("unsloth/llama-3-8b-Instruct")
correct_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
# Phi-3
template = phi3_template
correct_tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
correct_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
correct_tokenizer.chat_template = template
our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
assert(correct_prompt == our_prompt)
pass
def test_hf_gguf_equivalence(tokenizer, gguf_model = "./model-unsloth.F16.gguf"):
"""
Carefully checks the output of GGUF's tokenization and HF.
Can catch all tokenization bugs.
"""
import subprocess
import re
messages = [
{"role": "user", "content": "What is 2+2?"},
{"role": "assistant", "content": "It's 4."},
{"role": "user", "content": " But 2+2 is equal to 5. "},
{"role": "assistant", "content": "No I'm sure its 4."},
{"role": "user", "content": " No it's 100% 5! "},
]
prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
{}
### Input:
{}
### Response:
{}""".format(
"Describe the city given eloquently.", # instruction
"The lost city of Atlantis.", # input
"", # output - leave this blank for generation!
)
prompts = [ prompt, ]
if tokenizer.chat_template is not None:
prompt = tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True)
prompt = prompt.replace("'", "") # Subprocess does not like ''
prompt = remove_special_tokens(tokenizer, prompt)
prompts.append(prompt)
pass
for prompt in prompts:
command = f"./llama.cpp/llama-cli -m {gguf_model} -n 0 --temp 0.0 --verbose-prompt "\
f"--check-tensors -p '{prompt}'"
datas = []
with subprocess.Popen(command, shell = True, stdout = subprocess.PIPE, stderr = subprocess.STDOUT, bufsize = 1) as sp:
for line in sp.stdout:
datas.append(line.decode("utf-8", errors = "replace"))
pass
gguf_tokens = "".join(datas)
# Now extract GGUF tokenization attempt
gguf_tokenized = re.findall("([\d]{1,}) \-\> \'([^\']{1,})\'", gguf_tokens, flags = re.MULTILINE)
gguf_tokenized = [(int(x[0]), x[1],) for x in gguf_tokenized]
input_ids = tokenizer(prompt).input_ids
tokens = tokenizer.batch_decode(input_ids)
hf_tokenized = list(zip(input_ids, tokens))
# Compare to Huggingface
for j, (hf_token, gguf_token) in enumerate(zip(hf_tokenized, gguf_tokenized)):
if (hf_token[0] != gguf_token[0]):
print("Failed GGUF != HF at", j)
print("HF =", hf_token)
print("GGUF =", gguf_token)
print(hf_tokenized)
print()
print(gguf_tokenized)
print()
raise RuntimeError("Failed comparing GGUF to HF.")
pass
pass
return True
pass
build/lib/build/lib/unsloth/kernels/__init__.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .cross_entropy_loss import fast_cross_entropy_loss
from .rms_layernorm import fast_rms_layernorm
from .rope_embedding import fast_rope_embedding, inplace_rope_embedding
from .swiglu import swiglu_fg_kernel, swiglu_DWf_DW_dfg_kernel
from .geglu import (
geglu_exact_forward_kernel,
geglu_exact_backward_kernel,
geglu_approx_forward_kernel,
geglu_approx_backward_kernel,
)
from .fast_lora import (
get_lora_parameters,
get_lora_parameters_bias,
apply_lora_mlp_swiglu,
apply_lora_mlp_geglu_exact,
apply_lora_mlp_geglu_approx,
apply_lora_qkv,
apply_lora_o,
)
from .utils import fast_dequantize, fast_gemv, QUANT_STATE, fast_linear_forward, matmul_lora
from .flex_attention import HAS_FLEX_ATTENTION, slow_attention_softcapping
if HAS_FLEX_ATTENTION:
from .flex_attention import (
FLEX_ATTENTION_PADDING,
)
pass
try:
print("🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.")
except:
print("Unsloth: Will patch your computer to enable 2x faster free finetuning.")
pass
build/lib/build/lib/unsloth/kernels/cross_entropy_loss.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import triton
import triton.language as tl
import torch
from .utils import calculate_settings, MAX_FUSED_SIZE, triton_tanh
from transformers.models.llama.modeling_llama import logger
@triton.heuristics({"DO_SOFTCAPPING": lambda args: args["DO_SOFTCAPPING"],})
@triton.jit
def _cross_entropy_forward(
logits_ptr, logits_row_stride,
loss_ptr,
logsumexp_ptr,
labels_ptr,
VOCAB_SIZE : tl.constexpr,
BLOCK_SIZE : tl.constexpr,
DO_SOFTCAPPING : tl.constexpr,
SOFTCAP : tl.constexpr,
):
"""
Cross Entropy Loss = 1/n sum [ -yi log(Pi) ]
Pi = exp(xi) / sum(exp(xi))
CE_i = -y log(p) = -y log[ exp(x) / sum(exp(x)) ]
= -y [ x - log[sum(exp(x))] ]
= y * (log[sum(exp(x))] - x)
If y == 0: CE_i = 0
If y == 1: CE_i = logsumexp - x
logsumexp is also stable
Take y = log[sum(exp(x))]
exp(y) = sum(exp(x))
exp(y) = sum(exp(x - c)*exp(c)) Since e^(x-c)*e^c = e^x
exp(y) = exp(c)*sum(exp(x - c))
y = log(exp(c)*sum(exp(x - c)))
y = c + log[sum(exp(x - c))]
This means we can set c = max(x) to make sure
exp(x - c) always is exp(x - max(x)).
This ensures exp(x - max(x))'s maximum is 1 as exp(0) = 1.
"""
row_idx = tl.program_id(0)
logits_ptr += row_idx * logits_row_stride.to(tl.int64)
loss_ptr += row_idx
logsumexp_ptr += row_idx
labels_ptr += row_idx
col_offsets = tl.arange(0, BLOCK_SIZE)
mask = col_offsets < VOCAB_SIZE
label_idx = tl.load(labels_ptr).to(tl.int32)
logits = tl.load(logits_ptr + col_offsets, mask = mask, other = -float("inf"))
# Do logit softcapping for Gemma 2: t * tanh(1/t * x)
if DO_SOFTCAPPING: logits = SOFTCAP * triton_tanh(logits / SOFTCAP)
logits = logits.to(tl.float32)
c = tl.max(logits, 0)
logsumexp = c + tl.log(tl.sum(tl.exp(logits - c), 0))
if label_idx != -100:
x = tl.load(logits_ptr + label_idx)
# Do logit softcapping for Gemma 2: t * tanh(1/t * x)
if DO_SOFTCAPPING: x = SOFTCAP * triton_tanh(x / SOFTCAP)
loss = logsumexp - x.to(tl.float32)
else:
loss = 0.0
tl.store(logsumexp_ptr, logsumexp)
tl.store(loss_ptr, loss)
pass
@triton.heuristics({"DO_SOFTCAPPING": lambda args: args["DO_SOFTCAPPING"],})
@triton.jit
def _chunked_cross_entropy_forward(
logits_ptr, logits_row_stride,
loss_ptr,
logsumexp_ptr,
labels_ptr,
VOCAB_SIZE : tl.constexpr,
N_CHUNKS : tl.constexpr,
BLOCK_SIZE : tl.constexpr,
DO_SOFTCAPPING : tl.constexpr,
SOFTCAP : tl.constexpr,
):
"""
256K vocab divided in 4 chunks
|-65536-| |-65536-| |-65536-| |-65536-|
|-------| |-------| |-------| |-------|
|-------| |-------| |-------| |-------|
If y == 0: CE_i = 0
If y == 1: CE_i = logsumexp - x
Notice we can do logsumexp for each chunk and then
logsumexp[chunk_sum(logsumexp)] == logsumexp
chunk_sum = log[chunk_sum(logsumexp)]
= log[exp(logsumexp(a)) + ... + exp(logsumexp(z))]
= log[exp(log[sum(exp(a))]) + ... + exp(log[sum(exp(z))])]
= log[sum(exp(a)) + ... + sum(exp(z))]
= logsumexp(x)
This means we can perform a logsumexp for each chunk, then do a
final logsumexp reduction!
Ie do: logsumexp(chunked_logsumexp) - x
"""
row_idx = tl.program_id(0)
chunk_idx = tl.program_id(1)
logits_ptr += row_idx * logits_row_stride.to(tl.int64)
loss_ptr += row_idx
logsumexp_ptr += row_idx * N_CHUNKS + chunk_idx
labels_ptr += row_idx
col_offsets = chunk_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = col_offsets < VOCAB_SIZE
label_idx = tl.load(labels_ptr).to(tl.int32)
logits = tl.load(logits_ptr + col_offsets, mask = mask, other = -float("inf"))
# Do logit softcapping for Gemma 2: t * tanh(1/t * x)
if DO_SOFTCAPPING: logits = SOFTCAP * triton_tanh(logits / SOFTCAP)
logits = logits.to(tl.float32)
c = tl.max(logits, 0)
logsumexp = c + tl.log(tl.sum(tl.exp(logits - c), 0))
if chunk_idx == 0:
# logsumexp(chunked_logsumexp) - x
# Do the -x separately
if label_idx != -100:
x = tl.load(logits_ptr + label_idx).to(tl.float32)
# Do logit softcapping for Gemma 2: t * tanh(1/t * x)
if DO_SOFTCAPPING: x = SOFTCAP * triton_tanh(x / SOFTCAP)
loss = -1.0 * x.to(tl.float32)
else:
loss = 0.0
tl.store(loss_ptr, loss)
pass
tl.store(logsumexp_ptr, logsumexp)
pass
@triton.heuristics({"DO_SOFTCAPPING": lambda args: args["DO_SOFTCAPPING"],})
@triton.jit
def _cross_entropy_backward(
logits_ptr, logits_row_stride,
dloss_ptr, dloss_row_stride,
logsumexp_ptr,
labels_ptr,
VOCAB_SIZE : tl.constexpr,
BLOCK_SIZE : tl.constexpr,
DO_SOFTCAPPING : tl.constexpr,
SOFTCAP : tl.constexpr,
):
"""
CE_i = -y log(P) = y * (log[sum(exp(x))] - x)
dC/dx = d/dx (y * log[sum(exp(x))] - x * y)
From https://en.wikipedia.org/wiki/LogSumExp
d/dx logsumexp = exp(x) / sum(exp(x)) = softmax(x)
dC/dx = y * exp(x) / sum(exp(x)) - d/dx (x * y)
dC/dx = y * exp[ log[exp(x) / sum(exp(x))] ] using x = exp(log(x)) trick
dC/dx = y * exp[x - logsumexp] - d/dx (x * y)
If y == 0: dC/dx = 0
If y == 1 and x == label: dC/dlabel = exp[x - logsumexp] - 1
If y == 1 and x != label: dC/dx = exp[x - logsumexp]
"""
row_idx = tl.program_id(0)
block_idx = tl.program_id(1)
logits_ptr += row_idx * logits_row_stride.to(tl.int64)
dloss_ptr += row_idx * dloss_row_stride
col_offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = col_offsets < VOCAB_SIZE
label_idx = tl.load(labels_ptr + row_idx).to(tl.int32)
if label_idx != -100:
dloss = tl.load(dloss_ptr)
else:
dloss = 0.0
x = tl.load(logits_ptr + col_offsets, mask = mask, other = -float("inf"))
# Do logit softcapping for Gemma 2: t * tanh(1/t * x)
if DO_SOFTCAPPING:
# d/dx [t * tanh(1/t * x)] = 1 - tanh^2(1/t * x)
partial = triton_tanh(x / SOFTCAP)
x = SOFTCAP * partial
pass
logsumexp = tl.load(logsumexp_ptr + row_idx)
y = tl.exp(x.to(tl.float32) - logsumexp)
y = tl.where(
col_offsets == label_idx,
y - 1.0, # exp(x - logsumexp) - 1
y, # exp(x - logsumexp)
)
if DO_SOFTCAPPING:
# d/dx [t * tanh(1/t * x)] = 1 - tanh^2(1/t * x)
y = y * (1.0 - partial*partial)
pass
# If y == 0: dC/dx = 0 ==> we already masked it to be = 0, so dloss = 0.
tl.store(logits_ptr + col_offsets, dloss * y, mask = mask)
pass
MAX_FUSED_SIZE = 65536 # 2**16
class Fast_CrossEntropyLoss(torch.autograd.Function):
@staticmethod
def forward(ctx, logits, labels, logit_softcapping = 0):
n_rows, vocab_size = logits.shape
div, mod = divmod(vocab_size, MAX_FUSED_SIZE)
n_chunks = div + (mod != 0)
losses = torch.empty(n_rows, dtype = torch.float32, device = "cuda:0")
DO_SOFTCAPPING = (logit_softcapping != 0)
if n_chunks == 1:
# For small vocabs <= 65336 like Llama, Mistral
BLOCK_SIZE, num_warps = calculate_settings(vocab_size)
logsumexp = torch.empty(n_rows, dtype = torch.float32, device = "cuda:0")
_cross_entropy_forward[(n_rows,)](
logits, logits.stride(0),
losses,
logsumexp,
labels,
VOCAB_SIZE = vocab_size,
BLOCK_SIZE = BLOCK_SIZE,
DO_SOFTCAPPING = DO_SOFTCAPPING,
SOFTCAP = logit_softcapping,
num_warps = num_warps,
)
else:
# For large vocabs > 65336 like Gemma 256K
logsumexp = torch.empty((n_rows, n_chunks,), dtype = torch.float32, device = "cuda:0")
_chunked_cross_entropy_forward[(n_rows, n_chunks,)](
logits, logits.stride(0),
losses,
logsumexp,
labels,
VOCAB_SIZE = vocab_size,
N_CHUNKS = n_chunks,
BLOCK_SIZE = MAX_FUSED_SIZE,
DO_SOFTCAPPING = DO_SOFTCAPPING,
SOFTCAP = logit_softcapping,
num_warps = 32,
)
# logsumexp(chunked_logsumexp) - x
# Do the -x separately
logsumexp = torch.logsumexp(logsumexp, dim = 1) # Row sum
losses += logsumexp
losses.masked_fill_(labels == -100, 0) # Don't forget to mask padding out!
pass
ctx.save_for_backward(logits, logsumexp, labels)
ctx.DO_SOFTCAPPING = DO_SOFTCAPPING
ctx.logit_softcapping = logit_softcapping
return losses
pass
@staticmethod
def backward(ctx, dlosses):
logits, logsumexp, labels = ctx.saved_tensors
n_rows, vocab_size = logits.shape
BLOCK_SIZE = 4096
div, mod = divmod(vocab_size, BLOCK_SIZE)
n_blocks = div + (mod != 0)
_cross_entropy_backward[(n_rows, n_blocks,)](
logits, logits.stride(0),
dlosses, dlosses.stride(0),
logsumexp,
labels,
VOCAB_SIZE = vocab_size,
BLOCK_SIZE = BLOCK_SIZE,
DO_SOFTCAPPING = ctx.DO_SOFTCAPPING,
SOFTCAP = ctx.logit_softcapping,
num_warps = 8,
)
return logits, None, None,
pass
pass
def fast_cross_entropy_loss(logits, labels, logit_softcapping = 0):
"""
Arguments:
logits: (batch, seq_len, vocab_size)
labels: (batch, seq_len,)
Returns:
losses: float
"""
batch, seq_len, d = logits.shape
assert(labels.shape == (batch, seq_len))
loss = Fast_CrossEntropyLoss.apply(
logits.view(batch*seq_len, d),
labels.view(-1),
logit_softcapping,
)
n_items = torch.count_nonzero(labels != -100)
return loss.sum() / n_items
pass
build/lib/build/lib/unsloth/kernels/fast_lora.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from .utils import (
fast_dequantize,
QUANT_STATE,
get_lora_parameters,
get_lora_parameters_bias,
matmul_lora,
torch_amp_custom_fwd,
torch_amp_custom_bwd,
)
class LoRA_MLP(torch.autograd.Function):
"""
### LoRA weights
G = G + Ag @ Bg
U = U + Au @ Bu
W = W + Aw @ Bw
### SwiGLU(X)
e = X @ G
f = e * sigmoid(e)
g = X @ U
h = f * g
i = h @ W
### Backpropagation chain rule
See our blog post for more details
df = sigmoid(e) * (1 - f) + f
dC/dW = h.T @ dY
dC/dU = X.T @ (D @ W.T * f)
dC/dG = X.T @ (D @ W.T * df * g)
### Down projection LoRA weights
dC/dAw = dC/dW @ B.T
dC/dBw = A.T @ dC/dW
dC/dAw = h.T @ dY @ B.T
dC/dBw = A.T @ h.T @ dY
### Up projection LoRA weights
dC/dAu = X.T @ (D @ W.T * f) @ B.T
dC/dBu = A.T @ X.T @ (D @ W.T * f)
### Gate projection LoRA weights
dC/dAg = X.T @ (D @ W.T * df * g) @ B.T
dC/dBg = A.T @ X.T @ (D @ W.T * df * g)
Don't forget to see our blog post for more details!
"""
@staticmethod
@torch_amp_custom_fwd
def forward(ctx, X : torch.Tensor,
gateW, gateW_quant, gateA, gateB, gateS,
upW, upW_quant, upA, upB, upS,
downW, downW_quant, downA, downB, downS,
_forward_function, _backward_function,):
dtype = X.dtype
e = matmul_lora(X, gateW, gateW_quant, gateA, gateB, gateS)
g = matmul_lora(X, upW, upW_quant, upA, upB, upS)
h = _forward_function(e, g)
i = matmul_lora(h, downW, downW_quant, downA, downB, downS)
ctx.custom_saved_tensors = (
gateW, gateW_quant, gateS,
upW, upW_quant, upS,
downW, downW_quant, downS,
_backward_function,
)
ctx.save_for_backward(gateA, gateB, upA, upB, downA, downB,
X, e, g)
return i
pass
@staticmethod
@torch_amp_custom_bwd
def backward(ctx, dY : torch.Tensor):
gateW, gateW_quant, gateS, upW, upW_quant, upS, downW, downW_quant, downS, \
_backward_function = ctx.custom_saved_tensors
gateA, gateB, upA, upB, downA, downB, \
X, e, g = ctx.saved_tensors
gateA, gateB, upA, upB, downA, downB = \
gateA.t(), gateB.t(), upA.t(), upB.t(), downA.t(), downB.t()
batch, seq_len, hd = X.shape
dY = dY.view(-1, dY.shape[-1])
X = X .view(-1, X .shape[-1])
e = e .view(-1, e .shape[-1])
g = g .view(-1, g .shape[-1])
dtype = X.dtype
DW = matmul_lora(dY, downW.t(), downW_quant, downB, downA, downS)
DW, e, g = _backward_function(DW, e, g)
h, df, de = DW, e, g
# Down projection LoRA weights
d_downA = h.t() @ (dY @ downB.t())
d_downB = (downA.t() @ h.t()) @ dY
d_downA *= downS
d_downB *= downS
# Up projection LoRA weights
d_upA = X.t() @ (df @ upB.t())
d_upB = (upA.t() @ X.t()) @ df
d_upA *= upS
d_upB *= upS
# Gate projection LoRA weights
d_gateA = X.t() @ (de @ gateB.t())
d_gateB = (gateA.t() @ X.t()) @ de
d_gateA *= gateS
d_gateB *= gateS
# dX = matmul_lora(df, upW.t(), upW_quant, upB, upA, upS)
# dX += matmul_lora(de, gateW.t(), gateW_quant, gateB, gateA, gateS)
upW = fast_dequantize(upW.t(), upW_quant)
dX = torch.matmul(df, upW.t(), out = X)
del upW
dX += df @ upB.to(dtype).t() @ (upS * upA.to(dtype).t())
gateW = fast_dequantize(gateW.t(), gateW_quant)
dX += de @ gateW.t()
del gateW
dX += de @ gateB.to(dtype).t() @ (gateS * gateA.to(dtype).t())
# gateW, gateW_quant, gateA, gateB, gateS,
# upW, upW_quant, upA, upB, upS,
# downW, downW_quant, downA, downB, downS,
return dX.view(batch, seq_len, hd), \
None, None, d_gateA.t(), d_gateB.t(), None, \
None, None, d_upA.t(), d_upB.t(), None, \
None, None, d_downA.t(), d_downB.t(), None, \
None, None, # _backward and _forward
pass
pass
from .swiglu import swiglu_fg_kernel, swiglu_DWf_DW_dfg_kernel
def apply_lora_mlp_swiglu(self, X):
gateW, gateW_quant, gateA, gateB, gateS = get_lora_parameters(self.gate_proj)
upW, upW_quant, upA, upB, upS = get_lora_parameters(self. up_proj)
downW, downW_quant, downA, downB, downS = get_lora_parameters(self.down_proj)
out = LoRA_MLP.apply(X,
gateW, gateW_quant, gateA, gateB, gateS,
upW, upW_quant, upA, upB, upS,
downW, downW_quant, downA, downB, downS,
swiglu_fg_kernel, swiglu_DWf_DW_dfg_kernel,)
return out
pass
from .geglu import geglu_exact_forward_kernel, geglu_exact_backward_kernel
def apply_lora_mlp_geglu_exact(self, X):
gateW, gateW_quant, gateA, gateB, gateS = get_lora_parameters(self.gate_proj)
upW, upW_quant, upA, upB, upS = get_lora_parameters(self. up_proj)
downW, downW_quant, downA, downB, downS = get_lora_parameters(self.down_proj)
out = LoRA_MLP.apply(X,
gateW, gateW_quant, gateA, gateB, gateS,
upW, upW_quant, upA, upB, upS,
downW, downW_quant, downA, downB, downS,
geglu_exact_forward_kernel, geglu_exact_backward_kernel,)
return out
pass
from .geglu import geglu_approx_forward_kernel, geglu_approx_backward_kernel
def apply_lora_mlp_geglu_approx(self, X):
gateW, gateW_quant, gateA, gateB, gateS = get_lora_parameters(self.gate_proj)
upW, upW_quant, upA, upB, upS = get_lora_parameters(self. up_proj)
downW, downW_quant, downA, downB, downS = get_lora_parameters(self.down_proj)
out = LoRA_MLP.apply(X,
gateW, gateW_quant, gateA, gateB, gateS,
upW, upW_quant, upA, upB, upS,
downW, downW_quant, downA, downB, downS,
geglu_approx_forward_kernel, geglu_approx_backward_kernel,)
return out
pass
class LoRA_QKV(torch.autograd.Function):
"""
### LoRA weights
Wq = Wq + Aq @ Bq
Wk = Wk + Ak @ Bk
Wv = Wv + Av @ Bv
Q = X @ Wq = X @ Wq + X @ Aq @ Bq
K = X @ Wk = X @ Wk + X @ Ak @ Bk
V = X @ Wv = X @ Wv + X @ Av @ Bv
### Backpropagation chain rule
See our blogpost for more details.
dC/dWq = X.T @ D(Wq)
dC/dWk = X.T @ D(Wk)
dC/dWv = X.T @ D(Wv)
We then sum them all find dC/dX
### Q projection LoRA weights
dC/dAq = X.T @ D(Wq) @ B.T
dC/dBq = A.T @ X.T @ D(Wq)
### K projection LoRA weights
dC/dAk = X.T @ D(Wk) @ B.T
dC/dBk = A.T @ X.T @ D(Wk)
### V projection LoRA weights
dC/dAv = X.T @ D(Wv) @ B.T
dC/dBv = A.T @ X.T @ D(Wv)
"""
@staticmethod
@torch_amp_custom_fwd
def forward(ctx, X : torch.Tensor,
QW, QW_quant, QA, QB, QS,
KW, KW_quant, KA, KB, KS,
VW, VW_quant, VA, VB, VS,):
dtype = X.dtype
Q = matmul_lora(X, QW, QW_quant, QA, QB, QS)
K = matmul_lora(X, KW, KW_quant, KA, KB, KS)
V = matmul_lora(X, VW, VW_quant, VA, VB, VS)
ctx.custom_saved_tensors = (
QW, QW_quant, QS,
KW, KW_quant, KS,
VW, VW_quant, VS,
)
ctx.save_for_backward(X, QA, QB, KA, KB, VA, VB,)
return Q, K, V
pass
@staticmethod
@torch_amp_custom_bwd
def backward(ctx, dQ, dK, dV):
QW, QW_quant, QS, KW, KW_quant, KS, VW, VW_quant, VS = \
ctx.custom_saved_tensors
X, QA, QB, KA, KB, VA, VB, = ctx.saved_tensors
QA, QB, KA, KB, VA, VB = \
QA.t(), QB.t(), KA.t(), KB.t(), VA.t(), VB.t()
batch, seq_len, hd = X.shape
dQ = dQ.view(-1, dQ.shape[-1])
dK = dK.reshape(-1, dK.shape[-1]) # view doesn't work on K.T
dV = dV.view(-1, dV.shape[-1])
X = X .view(-1, X .shape[-1])
dtype = X.dtype
### Weight projection LoRA weights
# See our blogpost for more details.
# Q Projection
d_QA = X.t() @ (dQ @ QB.t())
d_QB = (QA.t() @ X.t()) @ dQ
d_QA *= QS
d_QB *= QS
# K Projection
d_KA = X.t() @ (dK @ KB.t())
d_KB = (KA.t() @ X.t()) @ dK
d_KA *= KS
d_KB *= KS
# V Projection
d_VA = X.t() @ (dV @ VB.t())
d_VB = (VA.t() @ X.t()) @ dV
d_VA *= VS
d_VB *= VS
# Combine derivatives to find dX
# dQ
QW = fast_dequantize(QW.t(), QW_quant)
dX = torch.matmul(dQ, QW.t(), out = X)
del QW
dX += (dQ @ QB.to(dtype).t() @ (QS * QA.to(dtype).t()))
# dK
KW = fast_dequantize(KW.t(), KW_quant)
dX += dK @ KW.t()
del KW
dX += dK @ KB.to(dtype).t() @ (KS * KA.to(dtype).t())
# dV
VW = fast_dequantize(VW.t(), VW_quant)
dX += dV @ VW.t()
del VW
dX += dV @ VB.to(dtype).t() @ (VS * VA.to(dtype).t())
# QW, QW_quant, QA, QB, QS,
# KW, KW_quant, KA, KB, KS,
# VW, VW_quant, VA, VB, VS,
return dX.view(batch, seq_len, hd), \
None, None, d_QA.t(), d_QB.t(), None, \
None, None, d_KA.t(), d_KB.t(), None, \
None, None, d_VA.t(), d_VB.t(), None
pass
pass
def apply_lora_qkv(self, X):
QW, QW_quant, QA, QB, QS = get_lora_parameters(self.q_proj)
KW, KW_quant, KA, KB, KS = get_lora_parameters(self.k_proj)
VW, VW_quant, VA, VB, VS = get_lora_parameters(self.v_proj)
Q, K, V = LoRA_QKV.apply(X,
QW, QW_quant, QA, QB, QS,
KW, KW_quant, KA, KB, KS,
VW, VW_quant, VA, VB, VS,
)
return Q, K, V
pass
class LoRA_W(torch.autograd.Function):
"""
### LoRA weights
Wq = Wq + Aq @ Bq
Wk = Wk + Ak @ Bk
Wv = Wv + Av @ Bv
Q = X @ Wq = X @ Wq + X @ Aq @ Bq
K = X @ Wk = X @ Wk + X @ Ak @ Bk
V = X @ Wv = X @ Wv + X @ Av @ Bv
### Backpropagation chain rule
dC/dWq = X.T @ D(Wq)
dC/dWk = X.T @ D(Wk)
dC/dWv = X.T @ D(Wv)
### Q projection LoRA weights
dC/dAq = X.T @ D(Wq) @ B.T
dC/dBq = A.T @ X.T @ D(Wq)
### K projection LoRA weights
dC/dAk = X.T @ D(Wk) @ B.T
dC/dBk = A.T @ X.T @ D(Wk)
### V projection LoRA weights
dC/dAv = X.T @ D(Wv) @ B.T
dC/dBv = A.T @ X.T @ D(Wv)
"""
@staticmethod
@torch_amp_custom_fwd
def forward(ctx, X : torch.Tensor,
W, W_quant, A, B, S):
dtype = X.dtype
XW = matmul_lora(X, W, W_quant, A, B, S)
ctx.custom_saved_tensors = (W, W_quant, S,)
ctx.save_for_backward(A, B, X)
return XW
pass
@staticmethod
@torch_amp_custom_bwd
def backward(ctx, dY : torch.Tensor):
W, W_quant, S = ctx.custom_saved_tensors
A, B, X = ctx.saved_tensors
A, B = A.t(), B.t()
batch, seq_len, hd = X.shape
dY = dY.reshape(-1, dY.shape[-1]) # Must be reshape
X = X .reshape(-1, X .shape[-1]) # Must be reshape
dtype = X.dtype
### Weight projection LoRA weights
# Weight projection
d_A = X.t() @ (dY @ B.t())
d_B = (A.t() @ X.t()) @ dY
d_A *= S
d_B *= S
# Get derivative for dX
W = fast_dequantize(W.t(), W_quant)
dX = dY @ W.t()
del W
dX += dY @ B.to(dtype).t() @ (S * A.to(dtype).t())
# W, W_quant, A, B, S
return dX.view(batch, seq_len, hd), \
None, None, d_A.t(), d_B.t(), None
pass
pass
def apply_lora_o(self, X):
OW, OW_quant, OA, OB, OS = get_lora_parameters(self.o_proj)
O = LoRA_W.apply(X, OW, OW_quant, OA, OB, OS)
return O
pass
build/lib/build/lib/unsloth/kernels/flex_attention.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from functools import lru_cache
from transformers.models.llama.modeling_llama import logger
torch_compile_options = {
"epilogue_fusion" : True,
"max_autotune" : True,
"shape_padding" : True,
"trace.enabled" : False, # Output Triton kernel outputs!
"triton.cudagraphs" : False,
}
# Flex Attention supported from torch 2.5 onwards only
import torch.nn
if hasattr(torch.nn, "attention"):
import torch.nn.attention
if hasattr(torch.nn.attention, "flex_attention"):
import torch.nn.attention.flex_attention
from torch.nn.attention.flex_attention import flex_attention
from torch.nn.attention.flex_attention import create_block_mask
FLEX_ATTENTION_PADDING = getattr(
torch.nn.attention.flex_attention,
"_DEFAULT_SPARSE_BLOCK_SIZE",
1,
)
flex_attention = torch.compile(flex_attention, dynamic = False)
HAS_FLEX_ATTENTION = True
else:
HAS_FLEX_ATTENTION = False
pass
else:
HAS_FLEX_ATTENTION = False
pass
# Logit softcapping
@torch.compile(fullgraph = True, dynamic = True, options = torch_compile_options)
def slow_attention_softcapping(Q, K, V, causal_mask, self, bsz, q_len):
n_heads = self.num_heads
head_dim = self.head_dim
n_kv_heads = self.num_key_value_heads
n_groups = self.num_key_value_groups
# Grouped query attention
K = K[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, q_len, head_dim)
V = V[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, q_len, head_dim)
K = K.reshape(bsz, n_heads, q_len, head_dim)
V = V.reshape(bsz, n_heads, q_len, head_dim)
# See https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
# Gemma 9b should use 256 and not 224 (hs / nah). 27b uses the below
# We default to using the config file itself
# s = self.config.hidden_size // self.config.num_attention_heads
s = self.config.query_pre_attn_scalar
t = self.config.attn_logit_softcapping
Q = Q * torch.tensor(s**-0.5, dtype = Q.dtype) # Follow Keras exactly
A = torch.matmul(Q, K.transpose(2, 3))
A = t * torch.tanh(A / t) # Logit softcapping
A += causal_mask[:q_len, :q_len]
# Much slower in torch compile!
# A.masked_fill_(causal_mask[:q_len, :q_len], -float("inf"))
A = torch.nn.functional.softmax(A, dim = -1, dtype = torch.float32).to(Q.dtype)
A = torch.matmul(A, V)
A = A.transpose(1, 2).contiguous()
A = A.reshape(bsz, q_len, n_heads*head_dim)
return A
pass
build/lib/build/lib/unsloth/kernels/geglu.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import triton
import triton.language as tl
import torch
from .utils import calculate_settings, triton_tanh
@triton.jit
def _exact_forward_kernel(e, g, h, n_elements, BLOCK_SIZE : tl.constexpr,):
block_idx = tl.program_id(0)
offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
# f = 1/2 * e * (1 + erf(1/sqrt(2) * e))
# h = f * up
e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
f_row = 0.5 * e_row * (tl.math.erf(tl.math.rsqrt(2.0) * e_row) + 1.0)
f_row = f_row.to(g_row.dtype) # Exact copy from HF
h_row = f_row * g_row
# Store h
tl.store(h + offsets, h_row, mask = mask)
pass
def geglu_exact_forward_kernel(gate, up):
batch, seq_len, hd = gate.shape
n_elements = gate.numel()
out = torch.empty((batch, seq_len, hd), dtype = gate.dtype, device = "cuda:0")
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
_exact_forward_kernel[grid](gate, up, out, n_elements, BLOCK_SIZE = 1024,)
return out
pass
@triton.jit
def _exact_backward_kernel(DW, e, g, n_elements, BLOCK_SIZE : tl.constexpr,):
"""
f = 1/2 * e * (1 + erf(1/sqrt(2) * e))
h = f * up
df/de (with help of Wolfram :)
df/de = 1/2 * (1 + erf(1/sqrt(2) * e)) + 1/sqrt(2*pi) * e * exp(-1/2 * e^2)
Reuse via
f = 1/2 * (1 + erf(1/sqrt(2) * e)) * e
"""
block_idx = tl.program_id(0)
offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
DW_row = tl.load(DW + offsets, mask = mask, other = 0)#.to(tl.float32)
e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
# Break e_row away for re-use
# f = 1/2 * e * (1 + erf(1/sqrt(2) * e))
f_partial_row = 0.5 * (tl.math.erf(tl.math.rsqrt(2.0) * e_row) + 1.0)
f_row = f_partial_row * e_row
f_row = f_row.to(DW_row.dtype)
# h = f * g
h_row = f_row * g_row
# df = DW * f
df_row = DW_row * f_row
# dg = DW * g
dg_row = DW_row * g_row
# df/de = 1/2 * (1 + erf(1/sqrt(2) * e)) + 1/sqrt(2*pi) * e * exp(-1/2 * e^2)
t = 0.3989422804014327 # 1/sqrt(2*pi)
df_de = f_partial_row + t * e_row * tl.exp(-0.5 * e_row * e_row)
de_row = dg_row.to(tl.float32) * df_de
de_row = de_row.to(DW_row.dtype)
# Store derivatives in buffers
tl.store(DW + offsets, h_row, mask = mask) # h = f * g
tl.store(e + offsets, df_row, mask = mask) # df = DW * f
tl.store(g + offsets, de_row, mask = mask) # de
pass
def geglu_exact_backward_kernel(DW, e, g):
batch_seq_len, hd = e.shape
n_elements = e.numel()
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
_exact_backward_kernel[grid](DW, e, g, n_elements, BLOCK_SIZE = 1024,)
return DW, e, g
pass
@triton.jit
def _approx_forward_kernel(e, g, h, n_elements, BLOCK_SIZE : tl.constexpr,):
block_idx = tl.program_id(0)
offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
# f = 1/2 * e * (1 + tanh( sqrt(2/pi) * (x + 0.044715 * x^3 ) ))
# f = 1/2 * e * (1 + tanh( sqrt(2/pi) * x * (1 + 0.044715 * x^2 ) ))
# h = f * up
s = 0.7978845608028654 # math.sqrt(2 / math.pi)
e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
f_row = 0.5 * e_row * (
triton_tanh(s * e_row * (1.0 + 0.044715 * e_row * e_row)) \
+ 1.0
)
f_row = f_row.to(g_row.dtype) # Exact copy from HF
h_row = f_row * g_row
# Store h
tl.store(h + offsets, h_row, mask = mask)
pass
def geglu_approx_forward_kernel(gate, up):
batch, seq_len, hd = gate.shape
n_elements = gate.numel()
out = torch.empty((batch, seq_len, hd), dtype = gate.dtype, device = "cuda:0")
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
_approx_forward_kernel[grid](gate, up, out, n_elements, BLOCK_SIZE = 1024,)
return out
pass
@triton.jit
def _approx_backward_kernel(DW, e, g, n_elements, BLOCK_SIZE : tl.constexpr,):
"""
f = 1/2 * e * (1 + tanh( sqrt(2/pi) * x * (1 + 0.044715 * x^2 ) ))
h = f * up
df/de (with help from https://arxiv.org/pdf/2305.12073.pdf :))
df/de = 1/2 * [1 + tanh( sqrt(2/pi) * x * (1 + 0.044715 * x^2 ) )] +
1/2 * sech^2 [ sqrt(2/pi) * x * (1 + 0.044715 * x^2 ) ] * \
( sqrt(2/pi) * x * (1 + 0.044715 * x^2 * 3 ) )
Notice sech^2(x) = 1 - tanh^2(x)
So reuse tanh( sqrt(2/pi) * x * (1 + 0.044715 * x^2 ) )
See https://www.desmos.com/calculator/nqprfoni6x
"""
block_idx = tl.program_id(0)
offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
DW_row = tl.load(DW + offsets, mask = mask, other = 0)#.to(tl.float32)
e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
# See https://www.desmos.com/calculator/nqprfoni6x
s = 0.7978845608028654 # math.sqrt(2 / math.pi)
a = s * e_row # a = sqrt(2 / pi) * x
b = a * 0.044715 * e_row * e_row # b = a * 0.044715 * x^2
T = 1.0 + triton_tanh(a + b)
T2 = 0.5 * T
# Q = 0.5 * -T * (T - 2.0) * (a + 3.0 * b)
Q2 = -T2 * (T - 2.0) * (a + 3.0 * b)
df_de = T2 + Q2 # 1/2 * (T + Q)
# f = 1/2 * e * (1 + tanh( sqrt(2/pi) * (x + 0.044715 * x^3 ) ))
f_row = T2 * e_row
f_row = f_row.to(DW_row.dtype)
# h = f * g
h_row = f_row * g_row
# df = DW * f
df_row = DW_row * f_row
# dg = DW * g
dg_row = DW_row * g_row
de_row = dg_row.to(tl.float32) * df_de
de_row = de_row.to(DW_row.dtype)
# Store derivatives in buffers
tl.store(DW + offsets, h_row, mask = mask) # h = f * g
tl.store(e + offsets, df_row, mask = mask) # df = DW * f
tl.store(g + offsets, de_row, mask = mask) # de
pass
def geglu_approx_backward_kernel(DW, e, g):
batch_seq_len, hd = e.shape
n_elements = e.numel()
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
_approx_backward_kernel[grid](DW, e, g, n_elements, BLOCK_SIZE = 1024,)
return DW, e, g
pass
build/lib/build/lib/unsloth/kernels/rms_layernorm.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import triton
import triton.language as tl
import torch
from .utils import calculate_settings
@triton.jit
def _rms_layernorm_forward(
Y, Y_row_stride,
X, X_row_stride,
W, W_row_stride,
r, r_row_stride,
n_cols, eps,
BLOCK_SIZE : tl.constexpr
):
"""
Fast RMS Layernorm kernel
Inspiration from a Triton tutorial:
https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
"""
row_idx = tl.program_id(0)
col_offsets = tl.arange(0, BLOCK_SIZE)
mask = col_offsets < n_cols
Y += row_idx * Y_row_stride
X += row_idx * X_row_stride
r += row_idx * r_row_stride
X_row = tl.load(X + col_offsets, mask = mask, other = 0).to(tl.float32)
W_row = tl.load(W + col_offsets, mask = mask, other = 0)#.to(tl.float32)
row_var = tl.sum(X_row * X_row, axis = 0) / n_cols
inv_var = tl.math.rsqrt(row_var + eps)
tl.store(r, inv_var)
normed = X_row * inv_var
normed = normed.to(W_row.dtype) # Exact copy from HF
output = normed * W_row
tl.store(Y + col_offsets, output, mask = mask)
pass
@triton.heuristics({"GEMMA": lambda args: args["GEMMA"],})
@triton.jit
def _rms_layernorm_backward(
dY, dY_row_stride,
X, X_row_stride,
W, W_row_stride,
r, r_row_stride,
dW, dW_row_stride,
n_cols, eps,
GEMMA : tl.constexpr,
BLOCK_SIZE : tl.constexpr,
):
"""
Fast RMS Layernorm kernel for the backward pass
Inspiration from a Triton tutorial:
https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
"""
row_idx = tl.program_id(0)
col_offsets = tl.arange(0, BLOCK_SIZE)
mask = col_offsets < n_cols
dY += row_idx * dY_row_stride
X += row_idx * X_row_stride
r += row_idx * r_row_stride
dY_row = tl.load(dY + col_offsets, mask = mask, other = 0).to(tl.float32)
X_row = tl.load(X + col_offsets, mask = mask, other = 0).to(tl.float32)
W_row = tl.load(W + col_offsets, mask = mask, other = 0).to(tl.float32)
# Get saved row variance
inv_var = tl.load(r).to(tl.float32)
normed = X_row * inv_var
if GEMMA: dY_W = dY_row * (W_row + 1.0)
else: dY_W = dY_row * W_row
rowsum_dY_normed = tl.sum(dY_W * normed, axis = 0)
output = inv_var/n_cols * (n_cols*dY_W - normed*rowsum_dY_normed)
tl.store(dY + col_offsets, output, mask = mask)
pass
@triton.jit
def _gemma_rms_layernorm_forward(
Y, Y_row_stride,
X, X_row_stride,
W, W_row_stride,
r, r_row_stride,
n_cols, eps,
BLOCK_SIZE : tl.constexpr,
):
# Copies https://github.com/google-deepmind/gemma/blob/main/gemma/layers.py#L31
# and https://github.com/keras-team/keras-nlp/blob/v0.8.2/keras_nlp/models/gemma/rms_normalization.py#L33
# exactly. Essentially all in float32!
row_idx = tl.program_id(0)
col_offsets = tl.arange(0, BLOCK_SIZE)
mask = col_offsets < n_cols
Y += row_idx * Y_row_stride
X += row_idx * X_row_stride
r += row_idx * r_row_stride
X_row = tl.load(X + col_offsets, mask = mask, other = 0).to(tl.float32)
W_row = tl.load(W + col_offsets, mask = mask, other = 0).to(tl.float32)
row_var = tl.sum(X_row * X_row, axis = 0) / n_cols
inv_var = tl.math.rsqrt(row_var + eps)
tl.store(r, inv_var)
normed = X_row * inv_var
output = normed * (W_row + 1.0)
tl.store(Y + col_offsets, output, mask = mask)
pass
class Fast_RMS_Layernorm(torch.autograd.Function):
@staticmethod
def forward(ctx, X, W, eps, gemma = False):
shape = X.shape
dim = shape[-1]
X = X.view(-1, dim)
n_rows, n_cols = X.shape
BLOCK_SIZE, num_warps = calculate_settings(n_cols)
Y = torch.empty((n_rows, n_cols), dtype = X.dtype, device = "cuda:0")
r = torch.empty(n_rows, dtype = torch.float32, device = "cuda:0")
fx = _gemma_rms_layernorm_forward if gemma else _rms_layernorm_forward
fx[(n_rows,)](
Y, Y.stride(0),
X, X.stride(0),
W, W.stride(0),
r, r.stride(0),
n_cols, eps,
BLOCK_SIZE = BLOCK_SIZE,
num_warps = num_warps,
)
ctx.eps = eps
ctx.BLOCK_SIZE = BLOCK_SIZE
ctx.num_warps = num_warps
ctx.GEMMA = gemma
ctx.save_for_backward(X, W, r)
return Y.view(*shape)
pass
@staticmethod
def backward(ctx, dY):
shape = dY.shape
dim = shape[-1]
dY = dY.view(-1, dim)
X, W, r = ctx.saved_tensors
n_rows, n_cols = dY.shape
dW = X
_rms_layernorm_backward[(n_rows,)](
dY, dY.stride(0),
X, X .stride(0),
W, W .stride(0),
r, r .stride(0),
dW, dW.stride(0),
n_cols, ctx.eps,
GEMMA = ctx.GEMMA,
BLOCK_SIZE = ctx.BLOCK_SIZE,
num_warps = ctx.num_warps,
)
dX = dY.view(*shape)
return dX, None, None, None
pass
pass
def fast_rms_layernorm(layernorm, X, gemma = False):
W = layernorm.weight
eps = layernorm.variance_epsilon
out = Fast_RMS_Layernorm.apply(X, W, eps, gemma)
return out
pass
build/lib/build/lib/unsloth/kernels/rope_embedding.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import triton
import triton.language as tl
import torch
from .utils import calculate_settings
ROPE_GROUP_SIZE = 4
@triton.heuristics({"BACKWARD_PASS": lambda args: args["BACKWARD_PASS"],})
@triton.jit
def _rope_embedding(
Q, Q_row_stride,
cos, cos_row_stride,
sin, sin_row_stride,
seqlen,
head_dim : tl.constexpr,
n_heads : tl.constexpr,
BACKWARD_PASS : tl.constexpr,
BLOCK_SIZE : tl.constexpr,
):
"""
Calculates the RoPE Embedding quickly
RoPE is Q * cos + rotate_half(Q) * sin
See our blog post for more info
"""
ROPE_GROUP_SIZE = 4
row_position = tl.program_id(0)
group_head_position = tl.program_id(1)
col_offsets = tl.arange(0, BLOCK_SIZE)
half_head_dim = head_dim // 2
mask = col_offsets < half_head_dim
sin1 = tl.load(sin + (row_position % seqlen)*sin_row_stride + \
half_head_dim*0 + col_offsets, mask = mask, other = 0)
cos1 = tl.load(cos + (row_position % seqlen)*cos_row_stride + \
half_head_dim*0 + col_offsets, mask = mask, other = 0)
if BACKWARD_PASS:
# See our blog post for more info.
sin1 = -sin1
pass
# [TODO] Autotune ROPE_GROUP_SIZE to be 1, 2, 4, 8
head_start = group_head_position * ROPE_GROUP_SIZE
head_end = min((head_start + ROPE_GROUP_SIZE), n_heads)
# 10% Faster kernel from [HuyNguyen-hust](https://github.com/unslothai/unsloth/pull/238)
for k in range(head_start, head_end):
offs_q1 = row_position * Q_row_stride + k * head_dim + col_offsets
offs_q2 = row_position * Q_row_stride + k * head_dim + col_offsets + half_head_dim
# For Gemma - sometimes RoPE must be done in float32 and not bfloat16
Q1 = tl.load(Q + offs_q1, mask = mask, other = 0).to(sin1.dtype)
Q2 = tl.load(Q + offs_q2, mask = mask, other = 0).to(sin1.dtype)
tl.store(Q + offs_q1, Q1*cos1 - Q2*sin1, mask = mask)
tl.store(Q + offs_q2, Q2*cos1 + Q1*sin1, mask = mask)
pass
pass
class Fast_RoPE_Embedding(torch.autograd.Function):
@staticmethod
def forward(ctx, Q, cos, sin):
cos, sin = cos.squeeze(), sin.squeeze()
batch, seq_len, n_heads, head_dim = Q.shape
Q = Q.view(batch*seq_len, n_heads*head_dim)
n_rows, n_cols = Q.shape
assert(seq_len <= cos.shape[0])
# [TODO] Changing blocksize to head_dim//2 seems to have
# some concurrency / un-deterministic issues.
BLOCK_SIZE, num_warps = calculate_settings(head_dim//2) # (head_dim//2)
# group_size = 4 # 4 or 8, too large group_size can hurt performance.
div, mod = divmod(n_heads, ROPE_GROUP_SIZE)
n_groups = div + (mod != 0)
_rope_embedding[(n_rows, n_groups, )](
Q, Q.stride(0),
cos, cos.stride(0),
sin, sin.stride(0),
seq_len,
head_dim, n_heads,
BACKWARD_PASS = False,
BLOCK_SIZE = BLOCK_SIZE,
num_warps = num_warps,
)
ctx.BLOCK_SIZE = BLOCK_SIZE
ctx.num_warps = num_warps
ctx.n_groups = n_groups
ctx.cos = cos
ctx.sin = sin
return Q.view(batch, seq_len, n_heads, head_dim)
pass
@staticmethod
def backward(ctx, dY):
batch, seq_len, n_heads, head_dim = dY.shape
dY = dY.reshape(batch*seq_len, n_heads*head_dim)
# Must be reshape not view
n_rows, n_cols = dY.shape
cos = ctx.cos
sin = ctx.sin
_rope_embedding[(n_rows, ctx.n_groups, )](
dY, dY .stride(0),
cos, cos.stride(0),
sin, sin.stride(0),
seq_len, head_dim, n_heads,
BACKWARD_PASS = True,
BLOCK_SIZE = ctx.BLOCK_SIZE,
num_warps = ctx.num_warps,
)
dY = dY.view(batch, seq_len, n_heads, head_dim)
return dY, None, None,
pass
pass
def fast_rope_embedding(Q, K, cos, sin):
Q = Fast_RoPE_Embedding.apply(Q.transpose(1, 2), cos, sin).transpose(1, 2)
K = Fast_RoPE_Embedding.apply(K.transpose(1, 2), cos, sin).transpose(1, 2)
return Q, K
pass
class Slow_RoPE_Embedding(torch.autograd.Function):
@staticmethod
def forward(ctx, Q, cos, sin, position_ids):
if position_ids is not None:
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
# Q * cos + rotate_half(Q) * sin
half = Q.shape[-1]//2
RH_Q = torch.cat((-Q[..., half:], Q[..., :half]), dim = -1)
Q *= cos
Q.addcmul_(RH_Q, sin)
# RH_Q *= sin
# Q += RH_Q
ctx.save_for_backward(cos, sin)
return Q
pass
@staticmethod
def backward(ctx, dY):
cos, sin = ctx.saved_tensors
# Q * cos + rotate_half.T(Q) * sin
half = dY.shape[-1]//2
RH_dY = torch.cat((dY[..., half:], -dY[..., :half]), dim = -1)
dY *= cos
dY.addcmul_(RH_dY, sin)
# RH_dY *= sin
# dY += RH_dY
return dY, None, None, None
pass
pass
def inplace_rope_embedding(Q, K, cos, sin, position_ids):
Q = Slow_RoPE_Embedding.apply(Q, cos, sin, position_ids)
K = Slow_RoPE_Embedding.apply(K, cos, sin, position_ids)
return Q, K
pass
build/lib/build/lib/unsloth/kernels/swiglu.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import triton
import triton.language as tl
import torch
from .utils import calculate_settings
@triton.jit
def _fg_kernel(e, g, h, n_elements, BLOCK_SIZE : tl.constexpr,):
block_idx = tl.program_id(0)
offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
# f = e * sigmoid(e)
f_row = e_row * tl.sigmoid(e_row) # e_row / (1 + tl.exp(-e_row))
f_row = f_row.to(g_row.dtype) # Exact copy from HF
# h = f * g
h_row = f_row * g_row
# Store h
tl.store(h + offsets, h_row, mask = mask)
pass
def swiglu_fg_kernel(e, g):
batch, seq_len, hd = e.shape
n_elements = e.numel()
h = torch.empty((batch, seq_len, hd), dtype = e.dtype, device = "cuda:0")
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
_fg_kernel[grid](e, g, h, n_elements, BLOCK_SIZE = 1024,)
return h
pass
@triton.jit
def _DWf_DW_dfg_kernel(DW, e, g, n_elements, BLOCK_SIZE : tl.constexpr,):
"""
e = e.float()
se = 1.0 / (1.0 + torch.exp(-e))
f = (se * e).to(dtype)
h = f * g
df = DW * f
dg = DW * g
de = (dg.float() * se * (1.0 + e * (1.0 - se))).to(dtype)
"""
block_idx = tl.program_id(0)
offsets = block_idx*BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
DW_row = tl.load(DW + offsets, mask = mask, other = 0)#.to(tl.float32)
e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
g_row = tl.load(g + offsets, mask = mask, other = 0)#.to(tl.float32)
# e = e.float()
# se = 1.0 / (1.0 + torch.exp(-e))
se_row = tl.sigmoid(e_row) # 1.0 / (1.0 + tl.exp(-e_row))
# f = (se * e).to(dtype)
f_row = se_row * e_row
f_row = f_row.to(DW_row.dtype)
# h = f * g
h_row = f_row * g_row
# df = DW * f
df_row = DW_row * f_row
# dg = DW * g
dg_row = DW_row * g_row
# de = (dg.float() * se * (1.0 + e * (1.0 - se))).to(dtype)
de_row = dg_row.to(tl.float32) * se_row * (1.0 + e_row * (1.0 - se_row))
de_row = de_row.to(DW_row.dtype)
# Store derivatives in buffers
tl.store(DW + offsets, h_row, mask = mask) # h = f * g
tl.store(e + offsets, df_row, mask = mask) # df = DW * f
tl.store(g + offsets, de_row, mask = mask) # de
pass
def swiglu_DWf_DW_dfg_kernel(DW, e, g):
batch_seq_len, hd = e.shape
n_elements = e.numel()
grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
_DWf_DW_dfg_kernel[grid](DW, e, g, n_elements, BLOCK_SIZE = 1024,)
return DW, e, g
pass
build/lib/build/lib/unsloth/kernels/utils.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import triton
MAX_FUSED_SIZE = 65536
next_power_of_2 = triton.next_power_of_2
# torch.cuda.amp.custom_fwd is deprecated >= 2.4
import torch
from packaging.version import Version
if Version(torch.__version__) < Version("2.4.0"):
torch_amp_custom_fwd = torch.cuda.amp.custom_fwd
torch_amp_custom_bwd = torch.cuda.amp.custom_bwd
else:
torch_amp_custom_fwd = torch.amp.custom_fwd(device_type = "cuda")
torch_amp_custom_bwd = torch.amp.custom_bwd(device_type = "cuda")
pass
# tl.math.tanh now is libdevice.tanh
from packaging.version import Version
import triton
if Version(triton.__version__) >= Version("3.0.0"):
from triton.language.extra import libdevice
triton_tanh = libdevice.tanh
else:
import triton.language as tl
triton_tanh = tl.math.tanh
pass
def calculate_settings(n):
BLOCK_SIZE = next_power_of_2(n)
if BLOCK_SIZE > MAX_FUSED_SIZE:
raise RuntimeError(f"Cannot launch Triton kernel since n = {n} exceeds "\
f"the maximum CUDA blocksize = {MAX_FUSED_SIZE}.")
num_warps = 4
if BLOCK_SIZE >= 32768: num_warps = 32
elif BLOCK_SIZE >= 8192: num_warps = 16
elif BLOCK_SIZE >= 2048: num_warps = 8
return BLOCK_SIZE, num_warps
pass
import bitsandbytes as bnb
get_ptr = bnb.functional.get_ptr
import ctypes
cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
cdequantize_blockwise_fp16_nf4 = bnb.functional.lib.cdequantize_blockwise_fp16_nf4
cdequantize_blockwise_bf16_nf4 = bnb.functional.lib.cdequantize_blockwise_bf16_nf4
cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemm_4bit_inference_naive_fp16
cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemm_4bit_inference_naive_bf16
def QUANT_STATE(W):
return getattr(W, "quant_state", None)
pass
def get_lora_parameters(proj):
# For DPO or disabled adapters
base_layer = (proj.base_layer if hasattr(proj, "base_layer") else proj)
W = base_layer.weight
if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
return W, QUANT_STATE(W), None, None, None
pass
active_adapter = proj.active_adapters[0] if \
hasattr(proj, "active_adapters") else proj.active_adapter
A = proj.lora_A [active_adapter].weight
B = proj.lora_B [active_adapter].weight
s = proj.scaling[active_adapter]
return W, QUANT_STATE(W), A, B, s
pass
def get_lora_parameters_bias(proj):
# For DPO or disabled adapters
base_layer = (proj.base_layer if hasattr(proj, "base_layer") else proj)
W = base_layer.weight
bias = base_layer.bias
if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
return W, QUANT_STATE(W), None, None, None, bias
pass
active_adapter = proj.active_adapters[0] if \
hasattr(proj, "active_adapters") else proj.active_adapter
A = proj.lora_A [active_adapter].weight
B = proj.lora_B [active_adapter].weight
s = proj.scaling[active_adapter]
return W, QUANT_STATE(W), A, B, s, bias
pass
def fast_dequantize(W, quant_state = None, out = None):
if quant_state is None: return W
if type(quant_state) is not list:
# New quant_state as a class
# https://github.com/TimDettmers/bitsandbytes/pull/763/files
absmax = quant_state.absmax
shape = quant_state.shape
dtype = quant_state.dtype
blocksize = quant_state.blocksize
offset = quant_state.offset
state2 = quant_state.state2
absmax2 = state2.absmax
code2 = state2.code
blocksize2 = state2.blocksize
else:
# Old quant_state as a list of lists
absmax, shape, dtype, blocksize, compressed_stats, _, _ = quant_state
offset, state2 = compressed_stats
absmax2, code2, blocksize2, _, _, _, _ = state2
pass
# Create weight matrix
if out is None:
out = torch.empty(shape, dtype = dtype, device = "cuda:0")
else:
assert(out.shape == shape)
assert(out.dtype == dtype)
# NF4 dequantization of statistics
n_elements_absmax = absmax.numel()
out_absmax = torch.empty(n_elements_absmax, dtype = torch.float32, device = "cuda:0")
# Do dequantization
ptr_out_absmax = get_ptr(out_absmax)
cdequantize_blockwise_fp32(
get_ptr(code2), get_ptr(absmax), get_ptr(absmax2), ptr_out_absmax,
ctypes.c_int(blocksize2), ctypes.c_int(n_elements_absmax)
)
out_absmax += offset
fx = cdequantize_blockwise_fp16_nf4 if dtype == torch.float16 else \
cdequantize_blockwise_bf16_nf4
fx(get_ptr(None), get_ptr(W), ptr_out_absmax, get_ptr(out),
ctypes.c_int(blocksize), ctypes.c_int(out.numel()))
# Careful returning transposed data
is_transposed = (True if W.shape[0] == 1 else False)
return out.t() if is_transposed else out
pass
def fast_gemv(X, W, quant_state, out = None):
if quant_state is None: return torch.matmul(X, W, out = out)
# For fast X @ W where seq_len == 1
# From https://github.com/TimDettmers/bitsandbytes/blob/main/bitsandbytes/functional.py#L1469
_, q_len, hd = X.shape
# assert(q_len == 1)
if type(quant_state) is not list:
# https://github.com/TimDettmers/bitsandbytes/pull/763/files
absmax = quant_state.absmax
shape = quant_state.shape
dtype = quant_state.dtype
blocksize = quant_state.blocksize
stats = quant_state.code
offset = quant_state.offset
state2 = quant_state.state2
absmax2 = state2.absmax
code2 = state2.code
blocksize2 = state2.blocksize
else:
absmax, shape, dtype, blocksize, compressed_stats, quant_type, stats = quant_state
offset, state2 = compressed_stats
absmax2, code2, blocksize2, _, _, _, _ = state2
pass
# assert(dtype == X.dtype)
bout = shape[0]
if out is None:
out = torch.empty((1, 1, bout,), dtype = dtype, device = "cuda:0")
# else:
# assert(out.shape == (1, 1, bout,))
# pass
n = 1
m = shape[0]
k = shape[1]
lda = shape[0]
ldc = shape[0]
ldb = (hd+1)//2
m = ctypes.c_int32(m)
n = ctypes.c_int32(n)
k = ctypes.c_int32(k)
lda = ctypes.c_int32(lda)
ldb = ctypes.c_int32(ldb)
ldc = ctypes.c_int32(ldc)
df = torch.empty(absmax.shape, dtype = torch.float32, device = "cuda:0")
cdequantize_blockwise_fp32(
get_ptr(code2), get_ptr(absmax), get_ptr(absmax2), get_ptr(df),
ctypes.c_int(blocksize2), ctypes.c_int(df.numel()),
)
df += offset
absmax = df
fx = cgemm_4bit_inference_naive_fp16 if dtype == torch.float16 else \
cgemm_4bit_inference_naive_bf16
blocksize = ctypes.c_int32(blocksize)
fx(m, n, k, get_ptr(X), get_ptr(W), get_ptr(absmax), get_ptr(stats), get_ptr(out),
lda, ldb, ldc, blocksize)
return out
pass
def fast_linear_forward(proj, X, temp_lora = None, out = None):
W, W_quant, lora_A, lora_B, lora_S, bias = get_lora_parameters_bias(proj)
bsz, q_len, in_dim = X.shape
if q_len != 1: return matmul_lora(X, W, W_quant, lora_A, lora_B, lora_S)
if W_quant is None:
out = torch.matmul(X, W.t(), out = out)
elif bsz == 1 and q_len == 1:
out = fast_gemv(X, W, W_quant, out = out)
else:
W = fast_dequantize(W.t(), W_quant)
out = torch.matmul(X, W, out = out)
pass
# Add in LoRA weights
if lora_A is not None:
out_dim = out.shape[2]
dtype = X.dtype
if not hasattr(lora_A, "_fast_lora"):
lora_A._fast_lora = lora_A.to(dtype)
lora_B._fast_lora = lora_B.to(dtype)
pass
if bsz == 1:
out = out.view(out_dim)
temp_lora = torch.mv(lora_A._fast_lora, X.ravel(), out = temp_lora)
out.addmv_(lora_B._fast_lora, temp_lora, alpha = lora_S)
else:
out = out.view(bsz, out_dim)
temp_lora = torch.mm(X.view(bsz, in_dim), lora_A._fast_lora.t(), out = temp_lora)
out.addmm_(temp_lora, lora_B._fast_lora.t(), alpha = lora_S)
pass
out = out.view(bsz, 1, out_dim)
pass
if bias is not None: out += bias
return out
pass
def matmul_lora(X, W, W_quant, A, B, s, out = None):
dtype = X.dtype
W = fast_dequantize(W.t(), W_quant)
if X.dim() == 3:
batch, seq_len, d = X.shape
X = X.view(-1, X.shape[-1])
reshape = True
else:
reshape = False
pass
out = torch.matmul(X, W, out = out)
if W_quant is not None: del W
if A is not None:
# LoRA is enabled
A, B = A.t(), B.t()
out += (X @ A.to(dtype)) @ (s * B.to(dtype))
pass
return out.view(batch, seq_len, -1) if reshape else out
pass
build/lib/build/lib/unsloth/models/__init__.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .loader import FastLanguageModel
from .llama import FastLlamaModel
from .mistral import FastMistralModel
from .qwen2 import FastQwen2Model
from .dpo import PatchDPOTrainer
from ._utils import is_bfloat16_supported
build/lib/build/lib/unsloth/models/_utils.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
__version__ = "2024.8"
__all__ = [
"prepare_model_for_kbit_training",
"xformers",
"xformers_attention",
"xformers_version",
"__version__",
"HAS_FLASH_ATTENTION",
"PRE_CHECK",
"platform_system",
"patch_tokenizer",
"get_statistics",
"Unsloth_Offloaded_Gradient_Checkpointer",
"offload_to_disk",
"offload_input_embeddings",
"offload_output_embeddings",
"is_bfloat16_supported",
"unsloth_offloaded_gradient_checkpoint",
"torch_compile_options",
"patch_linear_scaling",
"patch_llama_rope_scaling",
"check_nvidia",
"create_boolean_mask",
"torch_amp_custom_fwd",
"torch_amp_custom_bwd",
]
import torch
from typing import Union, Optional, List, Any, Callable, Tuple
from platform import system as platform_system
platform_system = platform_system()
import numpy as np
import warnings, subprocess, re, inspect, psutil, os, math
from packaging.version import Version
# =============================================
# Disable some warnings which can get annoying
warnings.filterwarnings(action = "ignore", category = UserWarning, module = "torch")
warnings.filterwarnings(action = "ignore", category = UserWarning, module = "huggingface_hub")
warnings.filterwarnings(action = "ignore", category = FutureWarning, module = "huggingface_hub")
warnings.filterwarnings(action = "ignore", category = RuntimeWarning, module = "subprocess")
warnings.filterwarnings(action = "ignore", category = UserWarning, module = "transformers")
warnings.filterwarnings(action = "ignore", category = FutureWarning, module = "accelerate")
warnings.filterwarnings(action = "ignore", category = RuntimeWarning, module = "multiprocessing")
warnings.filterwarnings(action = "ignore", category = RuntimeWarning, module = "multiprocess")
# Stop "Special tokens have been added in the vocabulary, ..."
import logging
logging.getLogger("transformers.tokenization_utils_base").setLevel(logging.CRITICAL+1)
# =============================================
# =============================================
# Edits all Config files to enable RoPE Scaling for all models
# Transformers had to update for Mistral Nemo 12b since Attention is (5120, 4096) now.
def patch_mistral_nemo_config(config):
if "head_dim (" not in config:
add_head_dim = "If it is not specified, will default to `8`.\n"\
" head_dim (`int`, *optional*, defaults to `hidden_size // num_attention_heads`):\n"\
" The attention head dimension."
config = config.replace("If it is not specified, will default to `8`.", add_head_dim)
add_head_dim = "num_key_value_heads=8,\n head_dim=None,"
config = config.replace("num_key_value_heads=8,", add_head_dim)
add_head_dim = "self.sliding_window = sliding_window\n self.head_dim = head_dim or hidden_size // num_attention_heads\n"
config = config.replace("self.sliding_window = sliding_window", add_head_dim)
pass
return config
pass
from transformers import __version__ as transformers_version
from transformers import PretrainedConfig
model_architectures = ["llama", "mistral", "gemma", "gemma2", "qwen2",]
for model_name in model_architectures:
config_filepath = f"transformers.models.{model_name}.configuration_{model_name}"
model_filepath = f"transformers.models.{model_name}.modeling_{model_name}"
config_filename = f"{model_name.title()}Config"
exec(f"from {config_filepath} import {config_filename}", globals())
try:
config = inspect.getsource(eval(config_filename))
except:
continue
if "rope_scaling" in config: continue
config = re.sub(
r"(\*\*kwargs)[\s]{0,}\,[\s]{0,}\)[\s]{0,}\:",
r"rope_scaling=None,"\
r"\n **kwargs):\n"\
r"\n self.rope_scaling = rope_scaling\n",
config,
)
# Just for Mistral Nemo
if model_name == "mistral":
if Version(transformers_version) <= Version("4.42.4"):
config = patch_mistral_nemo_config(config)
pass
exec(config, globals())
exec(f"import {config_filepath}", globals())
exec(f"{config_filepath}.{config_filename} = {config_filename}", globals())
pass
# =============================================
# =============================================
# torch.cuda.amp.custom_fwd is deprecated >= 2.4
import torch
if Version(torch.__version__) < Version("2.4.0"):
torch_amp_custom_fwd = torch.cuda.amp.custom_fwd
torch_amp_custom_bwd = torch.cuda.amp.custom_bwd
else:
torch_amp_custom_fwd = torch.amp.custom_fwd(device_type = "cuda")
torch_amp_custom_bwd = torch.amp.custom_bwd(device_type = "cuda")
pass
# =============================================
# =============================================
# Get Flash Attention v2 if Ampere (RTX 30xx, A100)
import bitsandbytes as bnb
from transformers.models.llama.modeling_llama import logger
from transformers import AutoTokenizer
major_version, minor_version = torch.cuda.get_device_capability()
SUPPORTS_BFLOAT16 = False
if major_version >= 8:
SUPPORTS_BFLOAT16 = True
try:
from flash_attn import flash_attn_func
# Check for CUDA linking errors "undefined symbol: _ZNK3c106SymIntltEl"
try:
from flash_attn.flash_attn_interface import flash_attn_cuda
HAS_FLASH_ATTENTION = True
except:
logger.warning_once(
"Unsloth: Your Flash Attention 2 installation seems to be broken?\n"\
"A possible explanation is you have a new CUDA version which isn't\n"\
"yet compatible with FA2? Please file a ticket to Unsloth or FA2.\n"\
"We shall now use Xformers instead, which gets a 0.01% performance hit.\n"\
"We found this negligible impact by benchmarking on 1x A100."
)
HAS_FLASH_ATTENTION = False
except:
HAS_FLASH_ATTENTION = False
else:
# Tri Dao's benchmark shows xformers is faster for now.
HAS_FLASH_ATTENTION = False
pass
import xformers.ops.fmha as xformers
xformers_attention = xformers.memory_efficient_attention
from xformers import __version__ as xformers_version
# Temporarily disable 0.0.27 and higher - inference issues
if Version(xformers_version) >= Version("0.0.27"):
raise ImportError(
"Unsloth: If you are in Colab, we updated the top cell install instructions - please change it to below "\
"then press Disconnect Runtime and then Restart it.\n"\
"\n"\
"%%capture\n"
"# Installs Unsloth, Xformers (Flash Attention) and all other packages!\n"
'!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"\n'
'!pip install --no-deps "xformers<0.0.27" "trl<0.9.0" peft accelerate bitsandbytes\n'\
'\n'\
f"Otherwise in local machines, your xformers version of {xformers_version} is too new.\n"\
'Please downgrade xformers via `pip install --force-reinstall "xformers<0.0.27"'
)
pass
# Check TRL version
from trl import __version__ as trl_version
if Version(xformers_version) >= Version("0.9.0"):
raise ImportError(
"Unsloth: If you are in Colab, we updated the top cell install instructions - please change it to below "\
"then press Disconnect Runtime and then Restart it.\n"\
"\n"\
"%%capture\n"
"# Installs Unsloth, Xformers (Flash Attention) and all other packages!\n"
'!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"\n'
'!pip install --no-deps "xformers<0.0.27" "trl<0.9.0" peft accelerate bitsandbytes\n'\
'\n'\
f"Otherwise in local machines, your TRL version of {trl_version} is too new.\n"\
'Please downgrade TRL via `pip install --force-reinstall "trl<0.9.0"'
)
pass
# =============================================
# =============================================
# Torch compile settings
# Just remove max_autotune_gemm warning
import functools
@functools.lru_cache(None)
def is_big_gpu(index):
sms = torch.cuda.get_device_properties(index).multi_processor_count
if sms < 80: # V100
# log.warning("not enough SMs to use max_autotune_gemm mode")
return False
return True
import torch._inductor.utils
torch._inductor.utils.is_big_gpu = is_big_gpu
# Torch compile arguments
torch_compile_arguments = [
"config.dce = True",
"config.memory_planning = True",
"config.memory_pool = 'combined'",
"config.coordinate_descent_tuning = True",
"config.max_autotune_gemm = False", # GEMM is unnecessary
"config.autotune_multi_device = False",
"config.max_autotune_gemm_backends = 'ATEN'", # Not much faster
"config.aggressive_fusion = False", # Careful changes results!
"config.cuda.enable_cuda_lto = True",
"config.cuda.use_fast_math = True",
"config.cuda.compile_opt_level = '-O2'",
]
# Torch dynamo arguments
torch_dynamo_arguments = [
"config.accumulated_cache_size_limit = 512", # Bump up a bit from 256
"config.suppress_errors = True", # Supress errors for now
"config.do_not_emit_runtime_asserts = True",
]
import torch._inductor.config as config
for _try_compile_argument in torch_compile_arguments:
try: exec(_try_compile_argument)
except: pass
pass
import torch._dynamo.config as config
for _try_dynamo_argument in torch_dynamo_arguments:
try: exec(_try_dynamo_argument)
except: pass
pass
torch_compile_options = {
"epilogue_fusion" : True,
"max_autotune" : True,
"shape_padding" : True,
"trace.enabled" : False, # Output Triton kernel outputs!
"triton.cudagraphs" : False,
}
# =============================================
def prepare_model_for_kbit_training(
model : Any,
use_gradient_checkpointing : Optional = True,
use_reentrant : Optional[bool] = True,
) -> Any:
"""
Calculates where to place the gradient checkpoints given n_layers.
We also freeze all other layers's gradients
Args:
model: Any LlamaModel with layers.
use_gradient_checkpointing (`bool`, *optional*):
Default enabled. Provides memory savings by not saving all activations,
but only some.
use_reentrant (`bool`, *optional*):
https://github.com/pytorch/pytorch/blob/main/torch/utils/checkpoint.py#L354
Optimal gradient checkpointing algorithm which will be the default in
future Pytorch versions.
"""
# Freeze all parameters except LoRA
import re
with torch.no_grad():
for name, param in model.named_parameters():
if ".lora_A." in name or ".lora_B." in name or ".lora_magnitude_vector" in name:
param.requires_grad_(True)
# Also must be in float32!
if param.dtype != torch.float32:
name = name.replace("base_model", "model", 1)
layer_number = re.search(r"\.[\d]{1,}\.", name).group(0)
name = name.replace(layer_number, f"[{layer_number[1:-1]}].")
name = name.replace(".weight", "", 1)
exec(f"{name}.to(torch.float32)")
pass
else:
param.requires_grad_(False)
pass
pass
# Gradient checkpointing!
if use_gradient_checkpointing == "unsloth":
# Saves VRAM!
original_model = model
while hasattr(original_model, "model"):
original_model._offloaded_gradient_checkpointing = True
original_model = original_model.model
pass
original_model._offloaded_gradient_checkpointing = True
model.gradient_checkpointing_enable()
elif use_gradient_checkpointing == True:
model.gradient_checkpointing_enable()
pass
# If use_reentrant = True which is the Pytorch default, we just make the input requires_grad.
if use_reentrant:
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
return model
pass
def patch_tokenizer(model, tokenizer):
"""
Phi3's pad_token isn't set. We set it to <|placeholder...
Llama-3 is <|reserved...
Llama-2 is <unk>
Check if pad_token is not the same as eos_token otherwise the loss will ignore it!!
Fixes https://github.com/unslothai/unsloth/issues/5
"""
possible_reserved_tokens = (
"<|reserved", # Llama-3
"<|placeholder", # Phi-3
"[control", # Mistral type models
"<pad>", # Mistral Nemo
"<|finetune_right_pad_id|>", # Llama-3.1
)
if model is not None:
model.config.update({"unsloth_version" : __version__})
bad_pad_token = False
if hasattr(tokenizer, "pad_token") and tokenizer.pad_token is not None:
# Check if pad_token is not the same as eos_token otherwise the loss will ignore it!!
bad_pad_token = tokenizer.eos_token == tokenizer.pad_token
elif hasattr(tokenizer, "pad_token") and tokenizer.pad_token is None:
bad_pad_token = True
else:
bad_pad_token = False
pass
if bad_pad_token:
# Find a better pad token
added_tokens = [str(x) for x in tokenizer.added_tokens_decoder.values()]
possible_pad_token = None
n_possible_pad_tokens = 0
for added_token in added_tokens[::-1]:
if added_token.startswith(possible_reserved_tokens):
if possible_pad_token is None: possible_pad_token = added_token
n_possible_pad_tokens += 1
# We must see at least 3 of the reserved tokens
if n_possible_pad_tokens >= 3: break
pass
pass
if n_possible_pad_tokens < 3: possible_pad_token = None
if possible_pad_token is None:
# Try unk_token
possible_pad_token = tokenizer.unk_token
pass
if possible_pad_token is None:
# Failure to find a good replacement!! We shall manually add one!
new_pad_token = "<|PAD_TOKEN|>"
while new_pad_token in tokenizer.get_vocab():
new_pad_token += "#"
pass
possible_pad_token = new_pad_token
pass
name = model.config._name_or_path if model is not None else "Model"
logger.warning_once(
f"{name} does not have a padding token! Will use pad_token = {possible_pad_token}."
)
# Edit pad_token
tokenizer.add_special_tokens({"pad_token" : possible_pad_token})
tokenizer.pad_token = possible_pad_token
if model is not None:
config = model.config.update({"pad_token_id" : tokenizer.pad_token_id})
else:
if model is not None:
if model.config.pad_token_id is None:
config = model.config.update({"pad_token_id" : tokenizer.pad_token_id})
return model, tokenizer
pass
# =============================================
# Weirdly LoraLayer.update_layer downcasts PEFT layers to float16??
# For mixed precision, we need it to be in float32 not float16.
from peft import __version__ as peft_version
if Version(peft_version) < Version("0.12.0"):
from peft.tuners.lora.layer import LoraLayer
import inspect, re
try:
source = inspect.getsource(LoraLayer.update_layer)
text = "if weight is not None:\n"
start = source.find(text) + len(text)
end = source.find("self.to(weight.device)", start)
spaces = re.findall(r"^([ ]{1,})break", source, flags = re.MULTILINE)[0]
source = source.replace(source[start : end], spaces)
spaces = len(re.match(r"[\s]{1,}", source).group(0))
lines = source.split("\n")
source = "\n".join(x[spaces:] for x in lines)
source = re.sub("([^\.])nn\.", r"\1torch.nn.", source)
source = source.replace("def update_layer", "def LoraLayer_update_layer")
exec(source, globals())
# Fix up incorrect downcasting of LoRA weights
from peft.tuners.lora.layer import LoraLayer
LoraLayer.update_layer = LoraLayer_update_layer
from peft.tuners.lora import LoraLayer
LoraLayer.update_layer = LoraLayer_update_layer
except:
logger.warning_once(
"Unsloth unsuccessfully patched LoraLayer.update_layer. Please file a bug report.\n"\
"Luckily, your training run will still work in the meantime!"
)
pass
pass
# =============================================
import psutil
def _get_statistics(statistics = None, force_download = True):
# We log some basic stats about which environment is being used.
# We simply download a README.md file from HF - all data is made public.
# This is simply so we can check if some envs are broken or not.
# You can disable this by commenting the below out
try:
n_cpus = psutil.cpu_count(logical = False)
keynames = "\n" + "\n".join(os.environ.keys())
if statistics is not None: pass
elif "\nCOLAB_" in keynames and n_cpus == 1: statistics = "colab"
elif "\nCOLAB_" in keynames: statistics = "colabpro"
elif "\nKAGGLE_" in keynames: statistics = "kaggle"
elif "\nRUNPOD_" in keynames: statistics = "runpod"
elif "\nAWS_" in keynames: statistics = "aws"
elif "\nAZURE_" in keynames: statistics = "azure"
elif "\nK_" in keynames or "\nFUNCTION_" in keynames: statistics = "gcp"
elif "\nINVOCATION_ID" in keynames: statistics = "lambda"
else: statistics = "other"
if statistics is not None:
from transformers import AutoModelForCausalLM
stats_model = AutoModelForCausalLM.from_pretrained(
f"unslothai/{statistics}",
force_download = force_download,
)
del stats_model
pass
except:
pass
pass
def get_statistics():
# We log some basic stats about which environment is being used.
# We simply download a README.md file from HF - all data is made public.
# This is simply so we can check if some envs are broken or not.
# You can disable this by commenting the below out
from huggingface_hub.utils import disable_progress_bars, enable_progress_bars, are_progress_bars_disabled
disabled = False
if not are_progress_bars_disabled():
disable_progress_bars()
disabled = True
pass
_get_statistics(None)
_get_statistics("repeat", force_download = False)
try:
vram = torch.cuda.get_device_properties(0).total_memory / 1024 / 1024 / 1024
if vram <= 8 : vram = 8
elif vram <= 16: vram = 16
elif vram <= 20: vram = 20
elif vram <= 24: vram = 24
elif vram <= 40: vram = 40
elif vram <= 48: vram = 48
elif vram <= 80: vram = 80
else: vram = 96
_get_statistics(f"vram-{vram}")
except:
pass
pass
try:
devices = torch.cuda.device_count()
_get_statistics(f"{devices if devices <= 8 else 9}")
except:
pass
if disabled: enable_progress_bars()
pass
def _calculate_n_gradient_checkpoints(
n_layers : int,
method : Optional[Union[str, int]] = "sqrt",
) -> List[int]:
assert(type(n_layers) is int and n_layers > 0)
if method is None: method = "sqrt"
if method == "sqrt":
n_checkpoints = int(n_layers**0.5)
elif type(method) is int and method > 0:
n_checkpoints = int(np.ceil(n_layers / method))
else:
raise ValueError("method must be 'sqrt' or an int >0 and <= n_layers.")
size = n_layers // n_checkpoints
sizes = np.full(n_checkpoints, size, dtype = int)
leftovers = n_layers % n_checkpoints
# We append leftovers from the right
for k in range(leftovers):
sizes[n_checkpoints-1-k] += 1
boundaries = np.hstack((0, np.cumsum(sizes)))
boundaries = boundaries.tolist()
return boundaries
pass
def calculate_n_gradient_checkpoints(
n_layers : int,
layers_per_checkpoint : Optional[Union[str, int]] = "sqrt",
) -> List[int]:
assert(type(n_layers) is int and n_layers > 0)
if layers_per_checkpoint is None or layers_per_checkpoint == 1:
return None
boundaries = _calculate_n_gradient_checkpoints(n_layers, layers_per_checkpoint)
assert(boundaries[0] == 0 and boundaries[-1] == n_layers)
assert(min(boundaries) == 0 and max(boundaries) == n_layers)
assert(np.diff(boundaries).min() >= 0)
return boundaries
pass
def prepare_n_gradient_checkpoints(
model : Any,
layers_per_checkpoint : Optional[Union[str, int]] = "sqrt",
use_reentrant : Optional[bool] = True,
) -> None:
"""
Calculates where to place the gradient checkpoints given n_layers.
Args:
model: Any LlamaModel with layers.
layers_per_checkpoint (`Union[str, int]`, *optional*):
Can either be `sqrt` or an integer for how many layers per checkpoint you want.
The more, the less memory usage, but can be slower. Default is `sqrt`.
Choose 1 for Pytorch gradient checkpointing. 2 to wrap 2 layers in 1 module etc.
use_reentrant (`bool`, *optional*):
https://github.com/pytorch/pytorch/blob/main/torch/utils/checkpoint.py#L354
Optimal gradient checkpointing algorithm `use_reentrant=False` which will
be the default in future Pytorch versions doesn't seem to work??
"""
_model = None
if hasattr(model, "layers"):
_model = model
elif hasattr(model, "model"):
if hasattr(model.model, "layers"):
_model = model.model
if _model is None:
raise TypeError("`model` or `model.model` does not have attribute `layers`. Are you sure this is a model?")
pass
if use_reentrant is False:
use_reentrant = True
pass
n_layers = len(_model.layers)
boundaries = calculate_n_gradient_checkpoints(n_layers, layers_per_checkpoint)
_model._gradient_checkpointing_boundaries = boundaries
_model._gradient_checkpointing_use_reentrant = use_reentrant
pass
class Unsloth_Offloaded_Gradient_Checkpointer(torch.autograd.Function):
"""
Saves VRAM by smartly offloading to RAM.
Tiny hit to performance, since we mask the movement via non blocking calls.
"""
@staticmethod
@torch_amp_custom_fwd
def forward(ctx, forward_function, hidden_states, *args):
saved_hidden_states = hidden_states.to("cpu", non_blocking = True)
with torch.no_grad():
output = forward_function(hidden_states, *args)
ctx.save_for_backward(saved_hidden_states)
ctx.forward_function = forward_function
ctx.args = args
return output
pass
@staticmethod
@torch_amp_custom_bwd
def backward(ctx, dY):
(hidden_states,) = ctx.saved_tensors
hidden_states = hidden_states.to("cuda:0", non_blocking = True).detach()
hidden_states.requires_grad = True
with torch.enable_grad():
(output,) = ctx.forward_function(hidden_states, *ctx.args)
torch.autograd.backward(output, dY)
return (None, hidden_states.grad,) + (None,)*len(ctx.args)
pass
pass
@torch._disable_dynamo
def unsloth_offloaded_gradient_checkpoint(function, *args, use_reentrant = None, **kwargs):
return Unsloth_Offloaded_Gradient_Checkpointer.apply(function, *args)
pass
# =============================================
# Fixes Bitsandbytes to remove missing warnings
from transformers.utils.quantization_config import BitsAndBytesConfig, QuantizationMethod
from inspect import getsource
from accelerate.utils.dataclasses import DistributedType
import re
BitsAndBytesConfig__init__ = getsource(BitsAndBytesConfig.__init__)
BitsAndBytesConfig__init__ = re.sub(
r"if[\s]{1,}kwargs\:[\s]{1,}.+?\n",
"",
BitsAndBytesConfig__init__,
flags = re.MULTILINE,
)
BitsAndBytesConfig__init__ = BitsAndBytesConfig__init__.split("\n")
length_spaces = len(re.match(r"[\s]{1,}", BitsAndBytesConfig__init__[0]).group(0))
BitsAndBytesConfig__init__ = "\n".join(x[length_spaces:] for x in BitsAndBytesConfig__init__)
BitsAndBytesConfig__init__ = BitsAndBytesConfig__init__.replace(
"__init__",
"_BitsAndBytesConfig__init__",
)
def _prepare_backend(
self, cpu: bool = False, sagemaker_dp = False, backend: str = None,
) -> tuple[str, DistributedType]:
return None, DistributedType.NO
pass
import accelerate.state
accelerate.state.PartialState._prepare_backend = _prepare_backend
import accelerate.accelerator
prepare = inspect.getsource(accelerate.accelerator.Accelerator.prepare)
prepare = prepare.split("\n")
spaces = prepare[0].find("def")
prepare = "\n".join(x[spaces:] for x in prepare)
x = "for obj in args:"
s = " "*spaces
prepare = prepare.replace(x, f'self.state.distributed_type = DistributedType.NO\n{s}{x}', 1)
exec(prepare, globals())
accelerate.accelerator.Accelerator.prepare = prepare
exec(BitsAndBytesConfig__init__, globals())
import transformers.utils.quantization_config
transformers.utils.quantization_config.BitsAndBytesConfig.__init__ = _BitsAndBytesConfig__init__
# =============================================
# Offloading to disk for modules (lm_head, embed_tokens)
import pickle
def offload_to_disk(W, model, name, temporary_location : str = "_unsloth_temporary_saved_buffers"):
file_location = os.path.join(temporary_location, model.config._name_or_path)
if not os.path.exists(file_location):
os.makedirs(file_location)
pass
filename = os.path.join(file_location, f"{name}.pt")
W = W.weight if hasattr(W, "weight") else W
torch.save(W, filename, pickle_module = pickle, pickle_protocol = pickle.HIGHEST_PROTOCOL,)
offloaded_W = torch.load(filename, map_location = "cpu", mmap = True)
offloaded_W._offloaded_file_location = filename
return offloaded_W
pass
def offload_input_embeddings(model, temporary_location : str = "_unsloth_temporary_saved_buffers"):
offloaded_W = offload_to_disk(model.get_input_embeddings(), model, "input_embeddings", temporary_location)
new_input_embeddings = torch.nn.Embedding.from_pretrained(offloaded_W)
new_input_embeddings._offloaded_file_location = offloaded_W._offloaded_file_location
model.set_input_embeddings(new_input_embeddings)
return
pass
def offload_output_embeddings(model, temporary_location : str = "_unsloth_temporary_saved_buffers"):
offloaded_W = offload_to_disk(model.get_output_embeddings(), model, "output_embeddings", temporary_location)
new_output_embeddings = torch.nn.Linear(1, 1, bias = None)
del new_output_embeddings.weight
new_output_embeddings.weight = offloaded_W
new_output_embeddings.in_features = offloaded_W.shape[1]
new_output_embeddings.out_features = offloaded_W.shape[0]
new_output_embeddings._offloaded_file_location = offloaded_W._offloaded_file_location
model.set_output_embeddings(new_output_embeddings)
return
pass
# Fixes a weird Torch 2.3 bug which says T4s have bfloat16
def is_bfloat16_supported():
return SUPPORTS_BFLOAT16
pass
# Patches models to add RoPE Scaling
def patch_linear_scaling(
model_name = "gemma2",
rope_module = None,
scaled_rope_module = None,
attention_module = None,
):
assert(rope_module is not None and scaled_rope_module is not None)
assert(attention_module is not None)
rope_name = rope_module.__name__
scaled_rope_name = scaled_rope_module.__name__
model_filepath = f"transformers.models.{model_name}.modeling_{model_name}"
exec_code = \
f"import torch.nn as nn\n"\
f"from typing import Union, Optional, List, Any, Callable, Tuple\n"\
f"from {model_filepath} import logger, "\
f"{model_name.title()}Attention, {model_name.title()}Config"
try:
function = inspect.getsource(attention_module.__init__)
except:
# Most likely already patched!
return None, None
where = function.find("def")
function = function.split("\n")
function = "\n".join(x[where:] for x in function)
init_name = f"{model_name.title()}Attention__init__"
function = function.replace("def __init__", f"def {init_name}")
function = function.replace(
"super().__init__()",
f"super({model_name.title()}Attention, self).__init__()",
)
fix_rope_function = """
if getattr(self.config, "rope_scaling", None) is None:
self.rotary_emb = {rope_function}(
dim = self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
else:
scaling_type = self.config.rope_scaling["type"]
scaling_factor = self.config.rope_scaling["factor"]
if scaling_type == "linear":
self.rotary_emb = {scaled_rope_function}(
dim = self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
else:
raise ValueError(f"Unknown RoPE scaling type {{scaling_type}}")
pass
"""
fix_rope_function = fix_rope_function.format(
rope_function = rope_module.__name__,
scaled_rope_function = scaled_rope_module.__name__,
)
rotary_emb = re.findall(
"self.rotary_emb = .+?\)", function,
flags = re.DOTALL | re.MULTILINE,
)
if len(rotary_emb) == 0: return None, function
rotary_emb = rotary_emb[0]
function = function.replace(rotary_emb, fix_rope_function, 1)
function = exec_code + "\n\n" + function
return init_name, function
pass
# Patches for Llama-3 LlamaExtendedRotaryEmbedding
def patch_llama_rope_scaling(
model_name = "llama",
rope_module = None,
scaled_rope_module = None,
extended_rope_module = None,
attention_module = None,
):
assert(\
rope_module is not None and \
scaled_rope_module is not None and \
extended_rope_module is not None
)
assert(attention_module is not None)
rope_name = rope_module.__name__
scaled_rope_name = scaled_rope_module.__name__
model_filepath = f"transformers.models.{model_name}.modeling_{model_name}"
exec_code = \
f"import torch.nn as nn\n"\
f"from typing import Union, Optional, List, Any, Callable, Tuple\n"\
f"from {model_filepath} import logger, "\
f"{model_name.title()}Attention, {model_name.title()}Config"
try:
function = inspect.getsource(attention_module.__init__)
except:
# Most likely already patched!
return None, None
where = function.find("def")
function = function.split("\n")
function = "\n".join(x[where:] for x in function)
init_name = f"{model_name.title()}Attention__init__"
function = function.replace("def __init__", f"def {init_name}")
function = function.replace(
"super().__init__()",
f"super({model_name.title()}Attention, self).__init__()",
)
fix_rope_function = """
if getattr(self.config, "rope_scaling", None) is None:
self.rotary_emb = {rope_function}(
dim = self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
else:
scaling_type1 = self.config.rope_scaling.get("type", None)
scaling_type2 = self.config.rope_scaling.get("rope_type", None)
scaling_type = scaling_type1 if scaling_type1 is not None else scaling_type2
scaling_factor = self.config.rope_scaling.get("factor")
if scaling_type == "linear":
self.rotary_emb = {scaled_rope_function}(
dim = self.head_dim,
max_position_embeddings=self.max_position_embeddings,
scaling_factor=scaling_factor,
base=self.rope_theta,
)
elif scaling_type == "llama3":
self.rotary_emb = {extended_rope_function}(
dim = self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
else:
raise ValueError(f"Unknown RoPE scaling type {{scaling_type}}")
pass
"""
fix_rope_function = fix_rope_function.format(
rope_function = rope_module.__name__,
scaled_rope_function = scaled_rope_module.__name__,
extended_rope_function = extended_rope_module.__name__,
)
rotary_emb = re.findall(
"self.rotary_emb = .+?\)", function,
flags = re.DOTALL | re.MULTILINE,
)
if len(rotary_emb) == 0: return None, function
rotary_emb = rotary_emb[0]
function = function.replace(rotary_emb, fix_rope_function, 1)
function = exec_code + "\n\n" + function
return init_name, function
pass
def check_nvidia():
# Unsloth doesn't work yet on AMD devices - we're working on it!
output = np.array([0,])
try:
output = subprocess.check_output("nvidia-smi --query-gpu=memory.used --format=csv", shell = True)
output = re.findall(rb'([\d]{1,})[\s]{1,}M', output)
output = np.array([int(x.decode('utf-8'))/1024 for x in output])
except:
if not torch.cuda.is_available():
raise RuntimeError("Unsloth: We do not support AMD / Intel machines yet - it is a work in progress!")
return output
pass
PRE_CHECK = check_nvidia()
def create_boolean_mask(n = 4096, sliding_window = 2048):
# Creates a boolean mask for attention
mask = torch.ones(n, n, dtype = torch.bool)
if sliding_window == 0:
return torch.triu(mask, diagonal = 1, out = mask)
pass
torch.triu(mask, diagonal = 0, out = mask)
torch.triu(mask.T, diagonal = -sliding_window, out = mask.T)
mask = mask.T
torch.logical_not(mask, out = mask)
return mask
pass
def test_mask_creation():
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
for n in range(2, 23):
for s in range(1, 23):
correct_mask = AttentionMaskConverter(
is_causal = True,
sliding_window = s,
).to_causal_4d(1, n, n, dtype = torch.float16,).squeeze(0).squeeze(0)
correct_mask = (correct_mask == correct_mask.min())
our_mask = create_boolean_mask(n = n, sliding_window = s)
assert(torch.all(correct_mask == our_mask))
pass
correct_mask = AttentionMaskConverter(
is_causal = True,
sliding_window = None,
).to_causal_4d(1, n, n, dtype = torch.float16,).squeeze(0).squeeze(0)
correct_mask = (correct_mask == correct_mask.min())
our_mask = create_boolean_mask(n = n, sliding_window = 0)
assert(torch.all(correct_mask == our_mask))
pass
pass
build/lib/build/lib/unsloth/models/dpo.py 0 → 100644
View file @ a380ace2
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
try:
from transformers.utils.notebook import (
IntervalStrategy,
NotebookTrainingTracker,
NotebookProgressCallback,
)
HAS_NOTEBOOK = True
except:
HAS_NOTEBOOK = False
pass
DPOTrainer_metrics = [
"rewards/chosen",
"rewards/rejected",
"rewards/accuracies",
"rewards/margins",
"logps/rejected",
"logps/chosen",
"logits/rejected",
"logits/chosen",
]
set_DPOTrainer_metrics = frozenset(DPOTrainer_metrics)
def NotebookProgressCallback_on_train_begin(self, args, state, control, **kwargs):
self.first_column = "Epoch" if args.evaluation_strategy == IntervalStrategy.EPOCH else "Step"
self.training_loss = 0
self.last_log = 0
column_names = [self.first_column] + ["Training Loss"]
if args.evaluation_strategy != IntervalStrategy.NO:
column_names.append("Validation Loss")
column_names += [x.replace("/", " / ") for x in DPOTrainer_metrics]
self.training_tracker = NotebookTrainingTracker(state.max_steps, column_names)
pass
def NotebookProgressCallback_on_log(self, args, state, control, logs=None, **kwargs):
# Only for when there is no evaluation
if args.evaluation_strategy == IntervalStrategy.NO and "loss" in logs:
values = {"Training Loss": logs["loss"]}
for metric in DPOTrainer_metrics:
values[metric.replace("/", " / ")] = logs[metric]
pass
# First column is necessarily Step since we're not in epoch eval strategy
values["Step"] = state.global_step
self.training_tracker.write_line(values)
pass
pass
def NotebookTrainingTracker_write_line(self, values):
"""
Write the values in the inner table.
Args:
values (`Dict[str, float]`): The values to display.
"""
if self.inner_table is None:
self.inner_table = [list(values.keys()), list(values.values())]
else:
columns = self.inner_table[0]
new_values = {}
for key, value in values.items():
lowered = key.lower()
if lowered in set_DPOTrainer_metrics:
new_values[lowered.replace("/", " / ")] = value
else:
new_values[key] = value
pass
values = new_values
self.inner_table[0] = columns
if len(self.inner_table) > 1:
last_values = self.inner_table[-1]
first_column = self.inner_table[0][0]
if last_values[0] != values[first_column]:
# write new line
self.inner_table.append([values[c] if c in values else "No Log" for c in columns])
else:
# update last line
new_values = values
for c in columns:
if c not in new_values.keys():
new_values[c] = last_values[columns.index(c)]
self.inner_table[-1] = [new_values[c] for c in columns]
else:
# Edit for evaluation purposes
self.inner_table.append([values[c] if c in values else 0 for c in columns])
pass
pass
pass
def PatchDPOTrainer():
if HAS_NOTEBOOK:
from transformers.trainer import is_in_notebook
if is_in_notebook():
# Patch DPO notebook printing
NotebookTrainingTracker.write_line = NotebookTrainingTracker_write_line
from transformers.trainer import DEFAULT_PROGRESS_CALLBACK
DEFAULT_PROGRESS_CALLBACK.on_train_begin = NotebookProgressCallback_on_train_begin
DEFAULT_PROGRESS_CALLBACK.on_log = NotebookProgressCallback_on_log
pass
pass
pass
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment