v1.0

src/llamafactory/model/model_utils/embedding.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 math
+from contextlib import nullcontext
+from typing import TYPE_CHECKING
+
+import torch
+from transformers.integrations import is_deepspeed_zero3_enabled
+
+from ...extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+
+logger = get_logger(__name__)
+
+
+def _noisy_mean_initialization(embed_weight: "torch.Tensor", num_new_tokens: int) -> None:
+    embedding_dim = embed_weight.size(1)
+    avg_weight = embed_weight[:-num_new_tokens].mean(dim=0, keepdim=True)
+    noise_weight = torch.empty_like(embed_weight[-num_new_tokens:])
+    noise_weight.normal_(mean=0, std=(1.0 / math.sqrt(embedding_dim)))
+    embed_weight[-num_new_tokens:] = avg_weight + noise_weight
+
+
+def resize_embedding_layer(model: "PreTrainedModel", tokenizer: "PreTrainedTokenizer") -> None:
+    r"""
+    Resize token embeddings.
+    """
+    if is_deepspeed_zero3_enabled():
+        import deepspeed  # type: ignore
+
+        params = [model.get_input_embeddings().weight]
+        if model.get_output_embeddings() is not None and not model.config.tie_word_embeddings:
+            params.append(model.get_output_embeddings().weight)
+
+        context_maybe_zero3 = deepspeed.zero.GatheredParameters(params, modifier_rank=0)
+    else:
+        context_maybe_zero3 = nullcontext()
+
+    with context_maybe_zero3:
+        current_embedding_size = model.get_input_embeddings().weight.size(0)
+
+    if len(tokenizer) > current_embedding_size:
+        if getattr(model, "quantization_method", None):
+            raise ValueError("Cannot resize embedding layers of a quantized model.")
+
+        if not isinstance(model.get_output_embeddings(), torch.nn.Linear):
+            raise ValueError("Current model does not support resizing embedding layers.")
+
+        model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=64)
+        with context_maybe_zero3:
+            new_embedding_size = model.get_input_embeddings().weight.size(0)
+            num_new_tokens = new_embedding_size - current_embedding_size
+            _noisy_mean_initialization(model.get_input_embeddings().weight.data, num_new_tokens)
+            _noisy_mean_initialization(model.get_output_embeddings().weight.data, num_new_tokens)
+
+        logger.info("Resized token embeddings from {} to {}.".format(current_embedding_size, new_embedding_size))

src/llamafactory/model/model_utils/liger_kernel.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 inspect
+from typing import TYPE_CHECKING
+
+from ...extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig
+
+    from ...hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def apply_liger_kernel(
+    config: "PretrainedConfig",
+    model_args: "ModelArguments",
+    is_trainable: bool,
+    require_logits: bool,
+) -> None:
+    if not is_trainable or not model_args.enable_liger_kernel:
+        return
+
+    model_type = getattr(config, "model_type", None)
+    if model_type == "gemma":
+        from liger_kernel.transformers import apply_liger_kernel_to_gemma as apply_liger_kernel
+    elif model_type == "gemma2":
+        from liger_kernel.transformers import apply_liger_kernel_to_gemma2 as apply_liger_kernel
+    elif model_type == "llama":
+        from liger_kernel.transformers import apply_liger_kernel_to_llama as apply_liger_kernel
+    elif model_type == "mistral":
+        from liger_kernel.transformers import apply_liger_kernel_to_mistral as apply_liger_kernel
+    elif model_type == "mixtral":
+        from liger_kernel.transformers import apply_liger_kernel_to_mixtral as apply_liger_kernel
+    elif model_type == "phi3":
+        from liger_kernel.transformers import apply_liger_kernel_to_phi3 as apply_liger_kernel
+    elif model_type == "qwen2":
+        from liger_kernel.transformers import apply_liger_kernel_to_qwen2 as apply_liger_kernel
+    elif model_type == "qwen2_vl":
+        from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl as apply_liger_kernel
+    else:
+        logger.warning("Current model does not support liger kernel.")
+        return
+
+    if require_logits and "fused_linear_cross_entropy" in inspect.signature(apply_liger_kernel).parameters:
+        logger.info("Current training stage does not support chunked cross entropy.")
+        kwargs = {"fused_linear_cross_entropy": False}
+    else:
+        kwargs = {}
+
+    apply_liger_kernel(**kwargs)
+    logger.info("Liger kernel has been applied to the model.")

src/llamafactory/model/model_utils/longlora.py

+# Copyright 2024 EleutherAI, HuggingFace Inc., Yukang Chen, and the LlamaFactory team.
+#
+# This code is based on the EleutherAI's GPT-NeoX and the HuggingFace's Transformers libraries.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llama/modeling_llama.py
+# This code is also inspired by the original LongLoRA implementation.
+# https://github.com/dvlab-research/LongLoRA/blob/main/llama_attn_replace.py
+#
+# 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 math
+from typing import TYPE_CHECKING, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from transformers.models.llama.modeling_llama import (
+    Cache,
+    LlamaAttention,
+    LlamaFlashAttention2,
+    LlamaSdpaAttention,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+from transformers.utils import logging
+from transformers.utils.versions import require_version
+
+from ...extras.constants import SUPPORTED_CLASS_FOR_S2ATTN
+from ...extras.logging import get_logger
+from ...extras.packages import is_transformers_version_greater_than_4_43
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig
+
+    from ...hparams import ModelArguments
+
+
+transformers_logger = logging.get_logger(__name__)
+
+
+# Modified from:
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llama/modeling_llama.py
+def llama_attention_forward(
+    self: "LlamaAttention",
+    hidden_states: "torch.Tensor",
+    attention_mask: Optional["torch.Tensor"] = None,
+    position_ids: Optional["torch.LongTensor"] = None,
+    past_key_value: Optional["Cache"] = None,
+    output_attentions: bool = False,
+    cache_position: Optional["torch.LongTensor"] = None,
+    position_embeddings: Optional[Tuple["torch.Tensor", "torch.Tensor"]] = None,
+    **kwargs,
+) -> Tuple["torch.Tensor", Optional["torch.Tensor"], Optional[Tuple["torch.Tensor"]]]:
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states: "torch.Tensor" = self.q_proj(hidden_states)
+    key_states: "torch.Tensor" = self.k_proj(hidden_states)
+    value_states: "torch.Tensor" = self.v_proj(hidden_states)
+
+    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+    value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+    if position_embeddings is None:
+        cos, sin = self.rotary_emb(value_states, position_ids)
+    else:
+        cos, sin = position_embeddings
+
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+    if past_key_value is not None:
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift
+        groupsz = int(q_len * getattr(self.config, "group_size_ratio"))
+        assert q_len % groupsz == 0, "q_len {} should be divisible by group size {}.".format(q_len, groupsz)
+        num_groups = q_len // groupsz
+
+        def shift(state: "torch.Tensor") -> "torch.Tensor":
+            state = state.transpose(1, 2)  # output: (bsz, seq_len, n_heads, head_dim)
+            state = torch.cat(
+                (state[:, :, : self.num_heads // 2], state[:, :, self.num_heads // 2 :].roll(-groupsz // 2, dims=1)),
+                dim=2,
+            )
+            return state.reshape(bsz * num_groups, groupsz, self.num_heads, self.head_dim).transpose(1, 2)
+
+        query_states, key_states, value_states = shift(query_states), shift(key_states), shift(value_states)
+        if attention_mask is not None:
+            attention_mask = attention_mask[:, :, :groupsz, :groupsz].repeat(num_groups, 1, 1, 1)
+
+    attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+    if attention_mask is not None:  # no matter the length, we just slice it
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    # upcast attention to fp32
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+    attn_output = torch.matmul(attn_weights, value_states)  # (bsz, :, seq_len, :) or (bsz * n_group, :, groupsz, :)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift back
+        attn_output.reshape(bsz, q_len, self.num_heads, self.head_dim)
+        attn_output = torch.cat(
+            (
+                attn_output[:, :, : self.num_heads // 2],
+                attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
+            ),
+            dim=2,
+        )
+
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output, attn_weights, past_key_value
+
+
+# Modified from:
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llama/modeling_llama.py
+def llama_flash_attention_2_forward(
+    self: "LlamaFlashAttention2",
+    hidden_states: "torch.Tensor",
+    attention_mask: Optional["torch.Tensor"] = None,
+    position_ids: Optional["torch.LongTensor"] = None,
+    past_key_value: Optional["Cache"] = None,
+    output_attentions: bool = False,
+    cache_position: Optional["torch.LongTensor"] = None,
+    position_embeddings: Optional[Tuple["torch.Tensor", "torch.Tensor"]] = None,
+    **kwargs,
+) -> Tuple["torch.Tensor", Optional["torch.Tensor"], Optional[Tuple["torch.Tensor"]]]:
+    # LlamaFlashAttention2 attention does not support output_attentions
+    output_attentions = False
+
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states: "torch.Tensor" = self.q_proj(hidden_states)
+    key_states: "torch.Tensor" = self.k_proj(hidden_states)
+    value_states: "torch.Tensor" = self.v_proj(hidden_states)
+
+    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+    value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+    if position_embeddings is None:
+        cos, sin = self.rotary_emb(value_states, position_ids)
+    else:
+        cos, sin = position_embeddings
+
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+    if past_key_value is not None:
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    # FlashAttention requires the input to have the shape (bsz, seq_len, n_heads, head_dim)
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    dropout_rate = self.attention_dropout if self.training else 0.0
+
+    input_dtype = query_states.dtype
+    if input_dtype == torch.float32:
+        if torch.is_autocast_enabled():
+            target_dtype = torch.get_autocast_gpu_dtype()
+        elif hasattr(self.config, "_pre_quantization_dtype"):
+            target_dtype = self.config._pre_quantization_dtype
+        else:
+            target_dtype = self.q_proj.weight.dtype
+
+        transformers_logger.warning_once("The input hidden states seems to be silently casted in float32.")
+        query_states = query_states.to(target_dtype)
+        key_states = key_states.to(target_dtype)
+        value_states = value_states.to(target_dtype)
+
+    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift
+        groupsz = int(q_len * getattr(self.config, "group_size_ratio"))
+        assert q_len % groupsz == 0, "q_len {} should be divisible by group size {}.".format(q_len, groupsz)
+        num_groups = q_len // groupsz
+
+        def shift(state: "torch.Tensor") -> "torch.Tensor":
+            state = torch.cat(
+                (state[:, :, : self.num_heads // 2], state[:, :, self.num_heads // 2 :].roll(-groupsz // 2, dims=1)),
+                dim=2,
+            )
+            return state.reshape(bsz * num_groups, groupsz, self.num_heads, self.head_dim)
+
+        query_states, key_states, value_states = shift(query_states), shift(key_states), shift(value_states)
+        if attention_mask is not None:
+            attention_mask = attention_mask[:, :groupsz].repeat(num_groups, 1)
+
+    if is_transformers_version_greater_than_4_43():
+        from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+        attn_output: "torch.Tensor" = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            query_states.size(1),
+            dropout=dropout_rate,
+            sliding_window=getattr(self, "sliding_window", None),
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=self.is_causal,
+        )
+    else:
+        attn_output: "torch.Tensor" = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, query_states.size(1), dropout=dropout_rate
+        )
+
+    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift back
+        attn_output.reshape(bsz, q_len, self.num_heads, self.head_dim)
+        attn_output = torch.cat(
+            (
+                attn_output[:, :, : self.num_heads // 2],
+                attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
+            ),
+            dim=2,
+        )
+
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output, attn_weights, past_key_value
+
+
+# Modified from:
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llama/modeling_llama.py
+def llama_sdpa_attention_forward(
+    self: "LlamaSdpaAttention",
+    hidden_states: "torch.Tensor",
+    attention_mask: Optional["torch.Tensor"] = None,
+    position_ids: Optional["torch.LongTensor"] = None,
+    past_key_value: Optional["Cache"] = None,
+    output_attentions: bool = False,
+    cache_position: Optional["torch.LongTensor"] = None,
+    position_embeddings: Optional[Tuple["torch.Tensor", "torch.Tensor"]] = None,
+    **kwargs,
+) -> Tuple["torch.Tensor", Optional["torch.Tensor"], Optional[Tuple["torch.Tensor"]]]:
+    if output_attentions:
+        transformers_logger.warning_once(
+            "SDPA does not support `output_attentions=True`. Falling back to the vanilla attention"
+        )
+        return llama_attention_forward(
+            self,
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states: "torch.Tensor" = self.q_proj(hidden_states)
+    key_states: "torch.Tensor" = self.k_proj(hidden_states)
+    value_states: "torch.Tensor" = self.v_proj(hidden_states)
+
+    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+    value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+    if position_embeddings is None:
+        cos, sin = self.rotary_emb(value_states, position_ids)
+    else:
+        cos, sin = position_embeddings
+
+    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+    if past_key_value is not None:
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift
+        groupsz = int(q_len * getattr(self.config, "group_size_ratio"))
+        assert q_len % groupsz == 0, "q_len {} should be divisible by group size {}.".format(q_len, groupsz)
+        num_groups = q_len // groupsz
+
+        def shift(state: "torch.Tensor") -> "torch.Tensor":
+            state = state.transpose(1, 2)  # output: (bsz, seq_len, n_heads, head_dim)
+            state = torch.cat(
+                (state[:, :, : self.num_heads // 2], state[:, :, self.num_heads // 2 :].roll(-groupsz // 2, dims=1)),
+                dim=2,
+            )
+            return state.reshape(bsz * num_groups, groupsz, self.num_heads, self.head_dim).transpose(1, 2)
+
+        query_states, key_states, value_states = shift(query_states), shift(key_states), shift(value_states)
+        if attention_mask is not None:
+            attention_mask = attention_mask[:, :, :groupsz, :groupsz].repeat(num_groups, 1, 1, 1)
+
+    causal_mask = attention_mask
+    if attention_mask is not None:
+        causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+    if query_states.device.type == "cuda" and causal_mask is not None:  # avoid pytorch bug
+        query_states = query_states.contiguous()
+        key_states = key_states.contiguous()
+        value_states = value_states.contiguous()
+
+    is_causal = True if causal_mask is None and q_len > 1 else False
+    attn_output = torch.nn.functional.scaled_dot_product_attention(
+        query_states,
+        key_states,
+        value_states,
+        attn_mask=causal_mask,
+        dropout_p=self.attention_dropout if self.training else 0.0,
+        is_causal=is_causal,
+    )
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift back
+        attn_output.reshape(bsz, q_len, self.num_heads, self.head_dim)
+        attn_output = torch.cat(
+            (
+                attn_output[:, :, : self.num_heads // 2],
+                attn_output[:, :, self.num_heads // 2 :].roll(groupsz // 2, dims=1),
+            ),
+            dim=2,
+        )
+
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+    attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value
+
+
+def _apply_llama_patch() -> None:
+    require_version("transformers>=4.41.2,<=4.45.2", "To fix: pip install transformers>=4.41.2,<=4.45.2")
+    LlamaAttention.forward = llama_attention_forward
+    LlamaFlashAttention2.forward = llama_flash_attention_2_forward
+    LlamaSdpaAttention.forward = llama_sdpa_attention_forward
+
+
+def configure_longlora(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
+    if not is_trainable or not model_args.shift_attn:
+        return
+
+    logger = get_logger(__name__)
+
+    if getattr(config, "model_type", None) in SUPPORTED_CLASS_FOR_S2ATTN:
+        setattr(config, "group_size_ratio", 0.25)
+        _apply_llama_patch()
+        logger.info("Using shift short attention with group_size_ratio=1/4.")
+    else:
+        logger.warning("Current model does not support shift short attention.")

src/llamafactory/model/model_utils/misc.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 typing import TYPE_CHECKING, List
+
+from ...extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedModel, PreTrainedTokenizer
+
+
+logger = get_logger(__name__)
+
+
+def find_all_linear_modules(model: "PreTrainedModel", freeze_vision_tower: bool) -> List[str]:
+    r"""
+    Finds all available modules to apply lora or galore.
+    """
+    model_type = getattr(model.config, "model_type", None)
+    forbidden_modules = {"lm_head"}
+    if model_type == "chatglm":
+        forbidden_modules.add("output_layer")
+    elif model_type == "internlm2":
+        forbidden_modules.add("output")
+    elif model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
+        forbidden_modules.add("multi_modal_projector")
+    elif model_type == "qwen2_vl":
+        forbidden_modules.add("merger")
+
+    if freeze_vision_tower:
+        if model_type == "qwen2_vl":
+            forbidden_modules.add("visual")
+        else:
+            forbidden_modules.add("vision_tower")
+
+    module_names = set()
+    for name, module in model.named_modules():
+        if any(forbidden_module in name for forbidden_module in forbidden_modules):
+            continue
+
+        if "Linear" in module.__class__.__name__ and "Embedding" not in module.__class__.__name__:
+            module_names.add(name.split(".")[-1])
+
+    logger.info("Found linear modules: {}".format(",".join(module_names)))
+    return list(module_names)
+
+
+def find_expanded_modules(model: "PreTrainedModel", target_modules: List[str], num_layer_trainable: int) -> List[str]:
+    r"""
+    Finds the modules in the expanded blocks to apply lora.
+    """
+    num_layers = getattr(model.config, "num_hidden_layers", None)
+    if not num_layers:
+        raise ValueError("Model was not supported.")
+
+    if num_layers % num_layer_trainable != 0:
+        raise ValueError(
+            "`num_layers` {} should be divisible by `num_layer_trainable` {}.".format(num_layers, num_layer_trainable)
+        )
+
+    stride = num_layers // num_layer_trainable
+    trainable_layer_ids = range(stride - 1, num_layers + stride - 1, stride)
+    trainable_layers = [".{:d}.".format(idx) for idx in trainable_layer_ids]
+    module_names = []
+    for name, _ in model.named_modules():
+        if any(target_module in name for target_module in target_modules) and any(
+            trainable_layer in name for trainable_layer in trainable_layers
+        ):
+            module_names.append(name)
+
+    logger.info("Apply lora to layers: {}".format(",".join(map(str, trainable_layer_ids))))
+    return module_names
+
+
+def register_autoclass(config: "PretrainedConfig", model: "PreTrainedModel", tokenizer: "PreTrainedTokenizer"):
+    if "AutoConfig" in getattr(config, "auto_map", {}):
+        config.__class__.register_for_auto_class()
+    if "AutoModelForCausalLM" in getattr(config, "auto_map", {}):
+        model.__class__.register_for_auto_class()
+    if "AutoTokenizer" in tokenizer.init_kwargs.get("auto_map", {}):
+        tokenizer.__class__.register_for_auto_class()

src/llamafactory/model/model_utils/mod.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 typing import TYPE_CHECKING
+
+from ...extras.constants import MOD_SUPPORTED_MODELS
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedModel
+
+    from ...hparams import ModelArguments
+
+
+def load_mod_pretrained_model(**init_kwargs) -> "PreTrainedModel":
+    from MoD import AutoMoDModelForCausalLM
+
+    return AutoMoDModelForCausalLM.from_pretrained(**init_kwargs)
+
+
+def convert_pretrained_model_to_mod(
+    model: "PreTrainedModel", config: "PretrainedConfig", model_args: "ModelArguments"
+) -> "PreTrainedModel":
+    from MoD import apply_mod_to_hf
+
+    if getattr(config, "model_type", None) not in MOD_SUPPORTED_MODELS:
+        raise ValueError("Current model is not supported by mixture-of-depth.")
+
+    model = apply_mod_to_hf(model)
+    model = model.to(model_args.compute_dtype)
+    return model

src/llamafactory/model/model_utils/moe.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 typing import TYPE_CHECKING, Sequence
+
+import torch
+from transformers.integrations import is_deepspeed_zero3_enabled
+from transformers.utils.versions import require_version
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedModel
+
+    from ...hparams import ModelArguments
+
+
+def _set_z3_leaf_modules(model: "PreTrainedModel", leaf_modules: Sequence["torch.nn.Module"]) -> None:
+    require_version("deepspeed>=0.13.0", "To fix: pip install deepspeed>=0.13.0")
+    from deepspeed.utils import set_z3_leaf_modules  # type: ignore
+
+    set_z3_leaf_modules(model, leaf_modules)
+
+
+def add_z3_leaf_module(model: "PreTrainedModel") -> None:
+    r"""
+    Sets module as a leaf module to skip partitioning in deepspeed zero3.
+    """
+    if not is_deepspeed_zero3_enabled():
+        return
+
+    model_type = getattr(model.config, "model_type", None)
+    if model_type == "dbrx":
+        from transformers.models.dbrx.modeling_dbrx import DbrxFFN
+
+        _set_z3_leaf_modules(model, [DbrxFFN])
+
+    if model_type == "jamba":
+        from transformers.models.jamba.modeling_jamba import JambaSparseMoeBlock
+
+        _set_z3_leaf_modules(model, [JambaSparseMoeBlock])
+
+    if model_type == "jetmoe":
+        from transformers.models.jetmoe.modeling_jetmoe import JetMoeMoA, JetMoeMoE
+
+        _set_z3_leaf_modules(model, [JetMoeMoA, JetMoeMoE])
+
+    if model_type == "mixtral":
+        from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
+
+        _set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
+
+    if model_type == "qwen2moe":
+        from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+        _set_z3_leaf_modules(model, [Qwen2MoeSparseMoeBlock])
+
+
+def configure_moe(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
+    model_type = getattr(config, "model_type", None)
+    if model_args.moe_aux_loss_coef is not None:
+        if model_type in ["jamba", "mixtral", "qwen2_moe"]:
+            setattr(config, "router_aux_loss_coef", model_args.moe_aux_loss_coef)
+
+        elif model_type == "deepseek":
+            setattr(config, "aux_loss_alpha", model_args.moe_aux_loss_coef)
+
+        elif model_type == "jetmoe":
+            setattr(config, "aux_loss_coef", model_args.moe_aux_loss_coef)
+
+    if model_type in ["dbrx", "jamba", "jetmoe", "mixtral", "qwen2_moe"]:
+        setattr(config, "output_router_logits", is_trainable)

src/llamafactory/model/model_utils/packing.py

+# Copyright 2024 Musab Gultekin and the LlamaFactory team.
+#
+# This code is based on the Musab Gultekin's functionary library.
+# https://github.com/MeetKai/functionary/blob/main/functionary/train/packing/monkey_patch_packing.py
+#
+# 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.
+#
+# MIT License
+#
+# Copyright (c) 2023 Musab Gultekin
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from typing import TYPE_CHECKING, Tuple
+
+import torch
+import torch.nn.functional as F
+from transformers.utils.versions import require_version
+
+from ...extras.constants import SUPPORTED_CLASS_FOR_BLOCK_DIAG_ATTN
+from ...extras.logging import get_logger
+from ...extras.packages import is_transformers_version_greater_than_4_43
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig
+
+    from ...hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def get_seqlens_in_batch(attention_mask: "torch.Tensor") -> "torch.Tensor":
+    r"""
+    Gets the sequnce lengths in the current batch.
+
+    e.g.
+    ```python
+    # input
+    [
+        [1, 1, 2, 2, 2, 0],
+        [1, 2, 2, 3, 3, 3],
+    ]
+    # output
+    [2, 3, 1, 2, 3]
+    ```
+    """
+    bsz = attention_mask.size(0)
+    dtype, device = attention_mask.dtype, attention_mask.device
+    max_num = torch.max(attention_mask).item()
+    counts: "torch.Tensor" = torch.zeros((bsz, max_num), dtype=dtype, device=device)
+    for i in range(max_num):
+        counts[:, i] = torch.sum(attention_mask == (i + 1), dim=-1)
+
+    counts = counts.flatten()
+    seqlens = counts[counts.nonzero().squeeze(dim=-1)]
+    return seqlens
+
+
+def get_unpad_data(attention_mask: "torch.Tensor") -> Tuple["torch.Tensor", "torch.Tensor", int]:
+    r"""
+    Prepares the indices and seqlens for flash attn varlen function.
+
+    Returns:
+        indices: indices of non-masked tokens from the flattened sequence.
+        cu_seqlens: the cumulative sequence lengths in the current batch, always starts from 0.
+        max_seqlen_in_batch: the largest seqlen in the current batch.
+
+    e.g.
+    ```python
+    # input
+    [
+        [1, 1, 2, 2, 2, 0],
+        [1, 2, 2, 3, 3, 3],
+    ]
+    # output
+    [0, 1, 2, 3, 4, 6, 7, 8, 9, 10, 11]
+    [0, 2, 5, 6, 8, 11]
+    3
+    ```
+    """
+    seqlens_in_batch = get_seqlens_in_batch(attention_mask)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return indices, cu_seqlens, max_seqlen_in_batch
+
+
+def _patch_for_block_diag_attn(model_type: str) -> None:
+    require_version("transformers>=4.41.2,<=4.45.2", "To fix: pip install transformers>=4.41.2,<=4.45.2")
+    if is_transformers_version_greater_than_4_43():
+        import transformers.modeling_flash_attention_utils
+
+        transformers.modeling_flash_attention_utils._get_unpad_data = get_unpad_data
+        return
+
+    import transformers.models
+
+    if model_type == "cohere":
+        transformers.models.cohere.modeling_cohere._get_unpad_data = get_unpad_data
+    elif model_type == "falcon":
+        transformers.models.falcon.modeling_falcon._get_unpad_data = get_unpad_data
+    elif model_type == "gemma":
+        transformers.models.gemma.modeling_gemma._get_unpad_data = get_unpad_data
+    elif model_type == "gemma2":
+        transformers.models.gemma2.modeling_gemma2._get_unpad_data = get_unpad_data
+    elif model_type == "llama":
+        transformers.models.llama.modeling_llama._get_unpad_data = get_unpad_data
+    elif model_type == "mistral":
+        transformers.models.mistral.modeling_mistral._get_unpad_data = get_unpad_data
+    elif model_type == "phi":
+        transformers.models.phi.modeling_phi._get_unpad_data = get_unpad_data
+    elif model_type == "phi3":
+        transformers.models.phi3.modeling_phi3._get_unpad_data = get_unpad_data
+    elif model_type == "qwen2":
+        transformers.models.qwen2.modeling_qwen2._get_unpad_data = get_unpad_data
+    elif model_type == "starcoder2":
+        transformers.models.starcoder2.modeling_starcoder2._get_unpad_data = get_unpad_data
+
+
+def configure_packing(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
+    if not is_trainable or not model_args.block_diag_attn:
+        return
+
+    model_type = getattr(config, "model_type", None)
+    if model_type in SUPPORTED_CLASS_FOR_BLOCK_DIAG_ATTN:
+        _patch_for_block_diag_attn(model_type)
+        logger.info("Using block diagonal attention for sequence packing without cross-attention.")
+    else:
+        raise ValueError("Current model does not support block diagonal attention.")

src/llamafactory/model/model_utils/quantization.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's Transformers and Optimum library.
+# https://github.com/huggingface/transformers/blob/v4.41.0/src/transformers/utils/quantization_config.py
+# https://github.com/huggingface/optimum/blob/v1.20.0/optimum/gptq/data.py
+#
+# 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 random
+from enum import Enum, unique
+from typing import TYPE_CHECKING, Any, Dict, List
+
+import torch
+from datasets import load_dataset
+from transformers import BitsAndBytesConfig, EetqConfig, GPTQConfig, HqqConfig
+from transformers.integrations import is_deepspeed_zero3_enabled
+from transformers.modeling_utils import is_fsdp_enabled
+from transformers.utils.versions import require_version
+
+from ...extras.constants import FILEEXT2TYPE
+from ...extras.logging import get_logger
+from ...extras.misc import get_current_device
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedTokenizer
+
+    from ...hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+@unique
+class QuantizationMethod(str, Enum):
+    r"""
+    Borrowed from `transformers.utils.quantization_config.QuantizationMethod`.
+    """
+
+    BITS_AND_BYTES = "bitsandbytes"
+    GPTQ = "gptq"
+    AWQ = "awq"
+    AQLM = "aqlm"
+    QUANTO = "quanto"
+    EETQ = "eetq"
+    HQQ = "hqq"
+
+
+def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "ModelArguments") -> List[Dict[str, Any]]:
+    r"""
+    Prepares the tokenized dataset to perform AutoGPTQ. Do not use tensor output for JSON serialization.
+    """
+    if os.path.isfile(model_args.export_quantization_dataset):
+        data_path = FILEEXT2TYPE.get(model_args.export_quantization_dataset.split(".")[-1], None)
+        data_files = model_args.export_quantization_dataset
+    else:
+        data_path = model_args.export_quantization_dataset
+        data_files = None
+
+    dataset = load_dataset(
+        path=data_path,
+        data_files=data_files,
+        split="train",
+        cache_dir=model_args.cache_dir,
+        token=model_args.hf_hub_token,
+    )
+
+    samples = []
+    maxlen = model_args.export_quantization_maxlen
+    for _ in range(model_args.export_quantization_nsamples):
+        n_try = 0
+        while True:
+            if n_try > 100:
+                raise ValueError("Cannot find satisfying example, considering decrease `export_quantization_maxlen`.")
+
+            sample_idx = random.randint(0, len(dataset) - 1)
+            sample: Dict[str, "torch.Tensor"] = tokenizer(dataset[sample_idx]["text"], return_tensors="pt")
+            n_try += 1
+            if sample["input_ids"].size(1) > maxlen:
+                break  # TODO: fix large maxlen
+
+        word_idx = random.randint(0, sample["input_ids"].size(1) - maxlen - 1)
+        input_ids = sample["input_ids"][:, word_idx : word_idx + maxlen]
+        attention_mask = sample["attention_mask"][:, word_idx : word_idx + maxlen]
+        samples.append({"input_ids": input_ids.tolist(), "attention_mask": attention_mask.tolist()})
+
+    return samples
+
+
+def configure_quantization(
+    config: "PretrainedConfig",
+    tokenizer: "PreTrainedTokenizer",
+    model_args: "ModelArguments",
+    init_kwargs: Dict[str, Any],
+) -> None:
+    r"""
+    Priority: PTQ-quantized (train/infer) > AutoGPTQ (export) > On-the-fly quantization (train/infer)
+    """
+    if getattr(config, "quantization_config", None):  # ptq
+        if model_args.quantization_bit is not None:
+            logger.warning("`quantization_bit` will not affect on the PTQ-quantized models.")
+
+        if is_deepspeed_zero3_enabled() or is_fsdp_enabled():
+            raise ValueError("DeepSpeed ZeRO-3 or FSDP is incompatible with PTQ-quantized models.")
+
+        quantization_config: Dict[str, Any] = getattr(config, "quantization_config", None)
+        quant_method = quantization_config.get("quant_method", "")
+
+        if quant_method == QuantizationMethod.GPTQ:
+            require_version("auto_gptq>=0.5.0", "To fix: pip install auto_gptq>=0.5.0")
+            quantization_config.pop("disable_exllama", None)  # remove deprecated args
+            quantization_config["use_exllama"] = False  # disable exllama
+
+        if quant_method == QuantizationMethod.AWQ:
+            require_version("autoawq", "To fix: pip install autoawq")
+
+        if quant_method == QuantizationMethod.AQLM:
+            require_version("aqlm>=1.1.0", "To fix: pip install aqlm[gpu]>=1.1.0")
+            quantization_config["bits"] = 2
+
+        quant_bits = quantization_config.get("bits", "?")
+        logger.info("Loading {}-bit {}-quantized model.".format(quant_bits, quant_method.upper()))
+
+    elif model_args.export_quantization_bit is not None:  # auto-gptq
+        if model_args.export_quantization_bit not in [8, 4, 3, 2]:
+            raise ValueError("AutoGPTQ only accepts 2/3/4/8-bit quantization.")
+
+        require_version("optimum>=1.17.0", "To fix: pip install optimum>=1.17.0")
+        require_version("auto_gptq>=0.5.0", "To fix: pip install auto_gptq>=0.5.0")
+        from accelerate.utils import get_max_memory
+
+        if getattr(config, "model_type", None) == "chatglm":
+            raise ValueError("ChatGLM model is not supported yet.")
+
+        init_kwargs["quantization_config"] = GPTQConfig(
+            bits=model_args.export_quantization_bit,
+            dataset=_get_quantization_dataset(tokenizer, model_args),
+        )
+        init_kwargs["device_map"] = "auto"
+        init_kwargs["max_memory"] = get_max_memory()
+        logger.info("Quantizing model to {} bit with AutoGPTQ.".format(model_args.export_quantization_bit))
+
+    elif model_args.quantization_bit is not None:  # on-the-fly
+        if model_args.quantization_method == QuantizationMethod.BITS_AND_BYTES.value:
+            if model_args.quantization_bit == 8:
+                require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
+                init_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_8bit=True)
+            elif model_args.quantization_bit == 4:
+                require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
+                init_kwargs["quantization_config"] = BitsAndBytesConfig(
+                    load_in_4bit=True,
+                    bnb_4bit_compute_dtype=model_args.compute_dtype,
+                    bnb_4bit_use_double_quant=model_args.double_quantization,
+                    bnb_4bit_quant_type=model_args.quantization_type,
+                    bnb_4bit_quant_storage=model_args.compute_dtype,  # crucial for fsdp+qlora
+                )
+            else:
+                raise ValueError("Bitsandbytes only accepts 4-bit or 8-bit quantization.")
+
+            # Do not assign device map if:
+            # 1. deepspeed zero3 or fsdp (train)
+            # 2. auto quantization device map (inference)
+            if is_deepspeed_zero3_enabled() or is_fsdp_enabled() or model_args.quantization_device_map == "auto":
+                if model_args.quantization_bit != 4:
+                    raise ValueError("Only 4-bit quantized model can use fsdp+qlora or auto device map.")
+
+                require_version("bitsandbytes>=0.43.0", "To fix: pip install bitsandbytes>=0.43.0")
+            else:
+                init_kwargs["device_map"] = {"": get_current_device()}  # change auto device map for inference
+
+            logger.info("Quantizing model to {} bit with bitsandbytes.".format(model_args.quantization_bit))
+        elif model_args.quantization_method == QuantizationMethod.HQQ.value:
+            if model_args.quantization_bit not in [8, 6, 5, 4, 3, 2, 1]:
+                raise ValueError("HQQ only accepts 1/2/3/4/5/6/8-bit quantization.")
+
+            if is_deepspeed_zero3_enabled() or is_fsdp_enabled():
+                raise ValueError("HQQ quantization is incompatible with DeepSpeed ZeRO-3 or FSDP.")
+
+            require_version("hqq", "To fix: pip install hqq")
+            init_kwargs["quantization_config"] = HqqConfig(
+                nbits=model_args.quantization_bit, quant_zero=False, quant_scale=False, axis=0
+            )  # use ATEN kernel (axis=0) for performance
+            logger.info("Quantizing model to {} bit with HQQ.".format(model_args.quantization_bit))
+        elif model_args.quantization_method == QuantizationMethod.EETQ.value:
+            if model_args.quantization_bit != 8:
+                raise ValueError("EETQ only accepts 8-bit quantization.")
+
+            if is_deepspeed_zero3_enabled() or is_fsdp_enabled():
+                raise ValueError("EETQ quantization is incompatible with DeepSpeed ZeRO-3 or FSDP.")
+
+            require_version("eetq", "To fix: pip install eetq")
+            init_kwargs["quantization_config"] = EetqConfig()
+            logger.info("Quantizing model to {} bit with EETQ.".format(model_args.quantization_bit))

src/llamafactory/model/model_utils/rope.py

+# Copyright 2024 LMSYS and the LlamaFactory team.
+# Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+# This code is inspired by the LMSYS's FastChat library.
+# https://github.com/lm-sys/FastChat/blob/v0.2.30/fastchat/train/train.py
+#
+# 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 math
+from typing import TYPE_CHECKING
+
+from ...extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig
+
+    from ...hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
+    if model_args.rope_scaling is None:
+        return
+
+    if not hasattr(config, "rope_scaling"):
+        logger.warning("Current model does not support RoPE scaling.")
+        return
+
+    if model_args.model_max_length is not None:
+        if is_trainable and model_args.rope_scaling == "dynamic":
+            logger.warning(
+                "Dynamic NTK scaling may not work well with fine-tuning. "
+                "See: https://github.com/huggingface/transformers/pull/24653"
+            )
+
+        current_max_length = getattr(config, "max_position_embeddings", None)
+        if current_max_length and model_args.model_max_length > current_max_length:
+            logger.info(
+                "Enlarge max model length from {} to {}.".format(current_max_length, model_args.model_max_length)
+            )
+            setattr(config, "max_position_embeddings", model_args.model_max_length)
+            scaling_factor = float(math.ceil(model_args.model_max_length / current_max_length))
+        else:
+            logger.warning("Input length is smaller than max length. Consider increase input length.")
+            scaling_factor = 1.0
+    else:
+        scaling_factor = 2.0
+
+    setattr(config, "rope_scaling", {"type": model_args.rope_scaling, "factor": scaling_factor})
+    logger.info(
+        "Using {} scaling strategy and setting scaling factor to {}".format(model_args.rope_scaling, scaling_factor)
+    )

src/llamafactory/model/model_utils/unsloth.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 typing import TYPE_CHECKING, Any, Dict, Optional
+
+from ...extras.logging import get_logger
+from ...extras.misc import get_current_device
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedModel
+
+    from ...hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def _get_unsloth_kwargs(
+    config: "PretrainedConfig", model_name_or_path: str, model_args: "ModelArguments"
+) -> Dict[str, Any]:
+    return {
+        "model_name": model_name_or_path,
+        "max_seq_length": model_args.model_max_length or 4096,
+        "dtype": model_args.compute_dtype,
+        "load_in_4bit": model_args.quantization_bit == 4,
+        "token": model_args.hf_hub_token,
+        "device_map": {"": get_current_device()},
+        "rope_scaling": getattr(config, "rope_scaling", None),
+        "fix_tokenizer": False,
+        "trust_remote_code": True,
+        "use_gradient_checkpointing": "unsloth",
+    }
+
+
+def load_unsloth_pretrained_model(
+    config: "PretrainedConfig", model_args: "ModelArguments"
+) -> Optional["PreTrainedModel"]:
+    r"""
+    Optionally loads pretrained model with unsloth. Used in training.
+    """
+    from unsloth import FastLanguageModel
+
+    unsloth_kwargs = _get_unsloth_kwargs(config, model_args.model_name_or_path, model_args)
+    try:
+        model, _ = FastLanguageModel.from_pretrained(**unsloth_kwargs)
+    except NotImplementedError:
+        logger.warning("Unsloth does not support model type {}.".format(getattr(config, "model_type", None)))
+        model = None
+        model_args.use_unsloth = False
+
+    return model
+
+
+def get_unsloth_peft_model(
+    model: "PreTrainedModel", model_args: "ModelArguments", peft_kwargs: Dict[str, Any]
+) -> "PreTrainedModel":
+    r"""
+    Gets the peft model for the pretrained model with unsloth. Used in training.
+    """
+    from unsloth import FastLanguageModel
+
+    unsloth_peft_kwargs = {
+        "model": model,
+        "max_seq_length": model_args.model_max_length,
+        "use_gradient_checkpointing": "unsloth",
+    }
+    return FastLanguageModel.get_peft_model(**peft_kwargs, **unsloth_peft_kwargs)
+
+
+def load_unsloth_peft_model(
+    config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool
+) -> "PreTrainedModel":
+    r"""
+    Loads peft model with unsloth. Used in both training and inference.
+    """
+    from unsloth import FastLanguageModel
+
+    unsloth_kwargs = _get_unsloth_kwargs(config, model_args.adapter_name_or_path[0], model_args)
+    try:
+        if not is_trainable:
+            unsloth_kwargs["use_gradient_checkpointing"] = False
+
+        model, _ = FastLanguageModel.from_pretrained(**unsloth_kwargs)
+    except NotImplementedError:
+        raise ValueError("Unsloth does not support model type {}.".format(getattr(config, "model_type", None)))
+
+    if not is_trainable:
+        FastLanguageModel.for_inference(model)
+
+    return model

src/llamafactory/model/model_utils/valuehead.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 typing import TYPE_CHECKING, Dict
+
+import torch
+from transformers.utils import cached_file
+
+from ...extras.constants import V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
+from ...extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+    from ...hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def load_valuehead_params(path_or_repo_id: str, model_args: "ModelArguments") -> Dict[str, torch.Tensor]:
+    r"""
+    Loads value head parameters from Hugging Face Hub or local disk.
+
+    Returns: dict with keys `v_head.summary.weight` and `v_head.summary.bias`.
+    """
+    kwargs = {"path_or_repo_id": path_or_repo_id, "cache_dir": model_args.cache_dir, "token": model_args.hf_hub_token}
+    err_text = ""
+
+    try:
+        from safetensors import safe_open
+
+        vhead_file = cached_file(filename=V_HEAD_SAFE_WEIGHTS_NAME, **kwargs)
+        with safe_open(vhead_file, framework="pt", device="cpu") as f:
+            return {key: f.get_tensor(key) for key in f.keys()}
+    except Exception as err:
+        err_text = str(err)
+
+    try:
+        vhead_file = cached_file(filename=V_HEAD_WEIGHTS_NAME, **kwargs)
+        return torch.load(vhead_file, map_location="cpu")
+    except Exception as err:
+        err_text = str(err)
+
+    logger.info("Provided path ({}) does not contain value head weights: {}.".format(path_or_repo_id, err_text))
+    logger.info("Ignore the above message if you are not resuming the training of a value head model.")
+    return None
+
+
+def prepare_valuehead_model(model: "PreTrainedModel") -> None:
+    if getattr(model.config, "model_type", None) == "llava":
+        setattr(model, "lm_head", model.language_model.get_output_embeddings())
+        setattr(model, "_keys_to_ignore_on_save", ["lm_head.weight"])
+
+    if getattr(model.config, "model_type", None) == "chatglm":
+        setattr(model, "lm_head", model.transformer.output_layer)
+        setattr(model, "_keys_to_ignore_on_save", ["lm_head.weight"])
+
+    if getattr(model.config, "model_type", None) == "internlm2":
+        setattr(model, "lm_head", model.output)
+        setattr(model, "_keys_to_ignore_on_save", ["lm_head.weight"])

src/llamafactory/model/model_utils/visual.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's Transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llava/modeling_llava.py
+#
+# 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 typing import TYPE_CHECKING, List, Sequence, Set, Tuple, Union
+
+import torch
+import transformers.models
+from transformers.activations import ACT2FN
+from transformers.utils import logging
+
+from ...extras.logging import get_logger
+
+
+if TYPE_CHECKING:
+    from transformers import LlavaConfig, PretrainedConfig, PreTrainedModel
+
+    from ...hparams import FinetuningArguments, ModelArguments
+
+
+logger = get_logger(__name__)
+transformers_logger = logging.get_logger(__name__)
+
+
+class LlavaMultiModalProjectorForYiVL(torch.nn.Module):
+    def __init__(self, config: "LlavaConfig") -> None:
+        super().__init__()
+
+        self.config = config
+        if config is None:
+            return
+
+        self.linear_1 = torch.nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
+        self.linear_2 = torch.nn.LayerNorm(config.text_config.hidden_size, bias=True)
+        self.linear_3 = torch.nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
+        self.linear_4 = torch.nn.LayerNorm(config.text_config.hidden_size, bias=True)
+        self.act = ACT2FN[config.projector_hidden_act]
+
+    def forward(self, image_features: "torch.Tensor") -> "torch.Tensor":
+        hidden_states = self.linear_1(image_features)
+        hidden_states = self.linear_2(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.linear_3(hidden_states)
+        hidden_states = self.linear_4(hidden_states)
+        if hidden_states.dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.linear_1.weight.dtype
+
+            transformers_logger.warning_once("The hidden states seems to be silently casted in float32.")
+            hidden_states = hidden_states.to(target_dtype)
+
+        return hidden_states
+
+
+class LlavaMultiModalProjectorForYiVLForVLLM(LlavaMultiModalProjectorForYiVL):
+    def __init__(self, vision_hidden_size: int, text_hidden_size: int, projector_hidden_act: str) -> None:
+        super().__init__(config=None)
+
+        self.linear_1 = torch.nn.Linear(vision_hidden_size, text_hidden_size, bias=True)
+        self.linear_2 = torch.nn.LayerNorm(text_hidden_size, bias=True)
+        self.linear_3 = torch.nn.Linear(text_hidden_size, text_hidden_size, bias=True)
+        self.linear_4 = torch.nn.LayerNorm(text_hidden_size, bias=True)
+        self.act = ACT2FN[projector_hidden_act]
+
+
+def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArguments") -> None:
+    r"""
+    Casts projector output to half precision for fine-tuning quantized VLMs.
+    """
+
+    def _mm_projector_forward_post_hook(
+        module: "torch.nn.Module", args: Tuple["torch.Tensor"], output: "torch.Tensor"
+    ) -> "torch.Tensor":
+        return output.to(model_args.compute_dtype)
+
+    if getattr(model, "quantization_method", None):
+        model_type = getattr(model.config, "model_type", None)
+        if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
+            mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
+        elif model_type == "qwen2_vl":
+            mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
+        else:
+            return
+
+        logger.info("Casting multimodal projector outputs in {}.".format(model_args.compute_dtype))
+        mm_projector.register_forward_hook(_mm_projector_forward_post_hook)
+
+
+def configure_visual_model(config: "PretrainedConfig") -> None:
+    r"""
+    Patches VLMs before loading them.
+    """
+    model_type = getattr(config, "model_type", None)
+    if model_type in [
+        "llava",
+        "llava_next",
+        "llava_next_video",
+        "paligemma",
+        "video_llava",
+    ]:  # required for ds zero3 and valuehead models
+        setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
+
+    if getattr(config, "is_yi_vl_derived_model", None):
+        logger.info("Detected Yi-VL model, applying projector patch.")
+        transformers.models.llava.modeling_llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVL
+
+
+def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "FinetuningArguments") -> Set[str]:
+    r"""
+    Freezes vision tower and language model for VLM full/freeze tuning.
+    """
+    model_type = getattr(config, "model_type", None)
+    forbidden_modules = set()
+    if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
+        if finetuning_args.freeze_vision_tower:
+            forbidden_modules.add("vision_tower")
+
+        if finetuning_args.train_mm_proj_only:
+            forbidden_modules.add("language_model")
+
+    elif model_type == "qwen2_vl":
+        if finetuning_args.freeze_vision_tower:
+            forbidden_modules.add("visual")
+
+        if finetuning_args.train_mm_proj_only:
+            raise ValueError("Qwen2-VL models do not support `train_mm_proj_only`.")
+
+    return forbidden_modules
+
+
+def get_image_seqlen(config: "PretrainedConfig") -> int:
+    r"""
+    Computes the number of special tokens per image.
+    """
+    model_type = getattr(config, "model_type", None)
+    if model_type == "llava":
+        image_seqlen = (config.vision_config.image_size // config.vision_config.patch_size) ** 2
+        if getattr(config, "vision_feature_select_strategy", "default") == "full":  # add [CLS] token
+            image_seqlen += 1
+    elif model_type == "paligemma":
+        image_seqlen = config.vision_config.num_image_tokens
+    else:
+        image_seqlen = -1
+
+    return image_seqlen
+
+
+def get_patch_size(config: "PretrainedConfig") -> int:
+    r"""
+    Computes the patch size of the vit.
+    """
+    patch_size = getattr(config.vision_config, "patch_size", -1)
+    return patch_size
+
+
+def get_vision_feature_select_strategy(config: "PretrainedConfig") -> int:
+    r"""
+    Get the vision_feature_select_strategy.
+    """
+    vision_feature_select_strategy = getattr(config, "vision_feature_select_strategy", "default")
+    return vision_feature_select_strategy
+
+
+def patch_target_modules(
+    config: "PretrainedConfig", finetuning_args: "FinetuningArguments", target_modules: Sequence[str]
+) -> Union[str, List[str]]:
+    r"""
+    Freezes vision tower for VLM LoRA tuning.
+    """
+    model_type = getattr(config, "model_type", None)
+    if finetuning_args.freeze_vision_tower:
+        if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
+            return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
+        elif model_type == "qwen2_vl":
+            return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))
+        else:
+            return target_modules
+    else:
+        if model_type == "qwen2_vl":
+            return "^(?!.*patch_embed).*(?:{}).*".format("|".join(target_modules))
+        else:
+            return target_modules

src/llamafactory/model/patcher.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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
+from types import MethodType
+from typing import TYPE_CHECKING, Any, Dict
+
+import torch
+from peft import PeftModel
+from transformers import PreTrainedModel, PreTrainedTokenizerBase, is_torch_npu_available
+from transformers.integrations import is_deepspeed_zero3_enabled
+from transformers.modeling_utils import is_fsdp_enabled
+
+from ..extras.logging import get_logger
+from ..extras.misc import infer_optim_dtype
+from .model_utils.attention import configure_attn_implementation, print_attn_implementation
+from .model_utils.checkpointing import prepare_model_for_training
+from .model_utils.embedding import resize_embedding_layer
+from .model_utils.longlora import configure_longlora
+from .model_utils.moe import add_z3_leaf_module, configure_moe
+from .model_utils.packing import configure_packing
+from .model_utils.quantization import configure_quantization
+from .model_utils.rope import configure_rope
+from .model_utils.valuehead import prepare_valuehead_model
+from .model_utils.visual import (
+    autocast_projector_dtype,
+    configure_visual_model,
+    get_image_seqlen,
+    get_patch_size,
+    get_vision_feature_select_strategy,
+)
+
+
+if TYPE_CHECKING:
+    from transformers import PretrainedConfig, PreTrainedTokenizer, ProcessorMixin
+    from trl import AutoModelForCausalLMWithValueHead
+
+    from ..hparams import ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def patch_tokenizer(tokenizer: "PreTrainedTokenizer") -> None:
+    if "PreTrainedTokenizerBase" not in str(tokenizer._pad.__func__):
+        tokenizer._pad = MethodType(PreTrainedTokenizerBase._pad, tokenizer)
+
+
+def patch_processor(
+    processor: "ProcessorMixin",
+    config: "PretrainedConfig",
+    tokenizer: "PreTrainedTokenizer",
+    model_args: "ModelArguments",
+) -> None:
+    setattr(processor, "tokenizer", tokenizer)
+    setattr(processor, "image_seqlen", get_image_seqlen(config))
+    setattr(processor, "image_resolution", model_args.image_resolution)
+    setattr(processor, "patch_size", get_patch_size(config))
+    setattr(processor, "video_resolution", model_args.video_resolution)
+    setattr(processor, "video_fps", model_args.video_fps)
+    setattr(processor, "video_maxlen", model_args.video_maxlen)
+    setattr(processor, "vision_feature_select_strategy", get_vision_feature_select_strategy(config))
+
+
+def patch_config(
+    config: "PretrainedConfig",
+    tokenizer: "PreTrainedTokenizer",
+    model_args: "ModelArguments",
+    init_kwargs: Dict[str, Any],
+    is_trainable: bool,
+) -> None:
+    if model_args.compute_dtype is None:  # priority: bf16 > fp16 > fp32
+        if model_args.infer_dtype != "auto" and not is_trainable:
+            model_args.compute_dtype = getattr(torch, model_args.infer_dtype)
+        else:
+            model_args.compute_dtype = infer_optim_dtype(model_dtype=getattr(config, "torch_dtype", None))
+
+    if is_torch_npu_available():
+        use_jit_compile = os.environ.get("JIT_COMPILE", "0").lower() in ["true", "1"]
+        torch.npu.set_compile_mode(jit_compile=use_jit_compile)
+
+    configure_attn_implementation(config, model_args, is_trainable)
+    configure_rope(config, model_args, is_trainable)
+    configure_longlora(config, model_args, is_trainable)
+    configure_quantization(config, tokenizer, model_args, init_kwargs)
+    configure_moe(config, model_args, is_trainable)
+    configure_visual_model(config)
+    configure_packing(config, model_args, is_trainable)
+
+    if model_args.use_cache and not is_trainable:
+        setattr(config, "use_cache", True)
+        logger.info("Using KV cache for faster generation.")
+
+    if getattr(config, "model_type", None) == "qwen":
+        setattr(config, "use_flash_attn", model_args.flash_attn == "fa2")
+        for dtype_name, dtype in [("fp16", torch.float16), ("bf16", torch.bfloat16), ("fp32", torch.float32)]:
+            setattr(config, dtype_name, model_args.compute_dtype == dtype)
+
+    if getattr(config, "model_type", None) == "qwen2" and is_trainable and model_args.flash_attn == "fa2":
+        setattr(config, "use_cache", False)  # qwen2 does not support use_cache when using flash attn
+
+    if "LlavaLlamaForCausalLM" in getattr(config, "architectures", []):
+        raise ValueError("Please download llava models with hf-compatible format: https://huggingface.co/llava-hf")
+
+    # deepspeed zero3 is not compatible with low_cpu_mem_usage
+    init_kwargs["low_cpu_mem_usage"] = model_args.low_cpu_mem_usage and (not is_deepspeed_zero3_enabled())
+
+    # cast data type of the model if:
+    # 1. not deepspeed zero3 and not fsdp (keep zero3 or fsdp in float32)
+    # 2. quantization_bit is not None (qlora)
+    if (not is_deepspeed_zero3_enabled() and not is_fsdp_enabled()) or model_args.quantization_bit is not None:
+        init_kwargs["torch_dtype"] = model_args.compute_dtype
+
+        if init_kwargs["low_cpu_mem_usage"]:  # device map requires low_cpu_mem_usage=True
+            if "device_map" not in init_kwargs and model_args.device_map:
+                init_kwargs["device_map"] = model_args.device_map
+
+            if init_kwargs.get("device_map", None) == "auto":
+                init_kwargs["offload_folder"] = model_args.offload_folder
+
+
+def patch_model(
+    model: "PreTrainedModel",
+    tokenizer: "PreTrainedTokenizer",
+    model_args: "ModelArguments",
+    is_trainable: bool,
+    add_valuehead: bool,
+) -> None:
+    gen_config = model.generation_config  # check and fix generation config
+    if not gen_config.do_sample and (
+        (gen_config.temperature is not None and gen_config.temperature != 1.0)
+        or (gen_config.top_p is not None and gen_config.top_p != 1.0)
+        or (gen_config.typical_p is not None and gen_config.typical_p != 1.0)
+    ):
+        gen_config.do_sample = True
+
+    if "GenerationMixin" not in str(model.generate.__func__):
+        model.generate = MethodType(PreTrainedModel.generate, model)
+
+    if add_valuehead:
+        prepare_valuehead_model(model)
+
+    if model_args.resize_vocab:
+        resize_embedding_layer(model, tokenizer)
+
+    if is_trainable:
+        prepare_model_for_training(model, model_args)
+        autocast_projector_dtype(model, model_args)
+        add_z3_leaf_module(model)
+
+    if not model_args.use_unsloth:
+        print_attn_implementation(model.config)
+
+    try:
+        model.add_model_tags(["llama-factory"])
+    except Exception:
+        logger.warning("Cannot properly tag the model.")
+
+
+def patch_valuehead_model(model: "AutoModelForCausalLMWithValueHead") -> None:
+    def tie_weights(self: "AutoModelForCausalLMWithValueHead") -> None:
+        if isinstance(self.pretrained_model, PreTrainedModel):
+            self.pretrained_model.tie_weights()
+
+    def get_input_embeddings(self: "AutoModelForCausalLMWithValueHead") -> torch.nn.Module:
+        if isinstance(self.pretrained_model, PreTrainedModel):
+            return self.pretrained_model.get_input_embeddings()
+
+    def get_output_embeddings(self: "AutoModelForCausalLMWithValueHead") -> torch.nn.Module:
+        if isinstance(self.pretrained_model, PreTrainedModel):
+            return self.pretrained_model.get_output_embeddings()
+
+    def create_or_update_model_card(self: "AutoModelForCausalLMWithValueHead", output_dir: str) -> None:
+        if isinstance(self.pretrained_model, PeftModel):
+            self.pretrained_model.create_or_update_model_card(output_dir)
+
+    ignore_modules = [name for name, _ in model.named_parameters() if "pretrained_model" in name]
+    setattr(model, "_keys_to_ignore_on_save", ignore_modules)
+    setattr(model, "tie_weights", MethodType(tie_weights, model))
+    setattr(model, "get_input_embeddings", MethodType(get_input_embeddings, model))
+    setattr(model, "get_output_embeddings", MethodType(get_output_embeddings, model))
+    setattr(model, "create_or_update_model_card", MethodType(create_or_update_model_card, model))

src/llamafactory/train/callbacks.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 json
+import logging
+import os
+import signal
+import sys
+import time
+from concurrent.futures import ThreadPoolExecutor
+from datetime import timedelta
+from typing import TYPE_CHECKING, Any, Dict, Optional
+
+import torch
+import transformers
+from peft import PeftModel
+from transformers import PreTrainedModel, ProcessorMixin, TrainerCallback
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length
+from transformers.utils import (
+    SAFE_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    is_safetensors_available,
+)
+from typing_extensions import override
+
+from ..extras.constants import TRAINER_LOG, V_HEAD_SAFE_WEIGHTS_NAME, V_HEAD_WEIGHTS_NAME
+from ..extras.logging import LoggerHandler, get_logger
+from ..extras.misc import get_peak_memory
+
+
+if is_safetensors_available():
+    from safetensors import safe_open
+    from safetensors.torch import save_file
+
+if TYPE_CHECKING:
+    from transformers import TrainerControl, TrainerState, TrainingArguments
+    from trl import AutoModelForCausalLMWithValueHead
+
+
+logger = get_logger(__name__)
+
+
+def fix_valuehead_checkpoint(
+    model: "AutoModelForCausalLMWithValueHead", output_dir: str, safe_serialization: bool
+) -> None:
+    r"""
+    The model is already unwrapped.
+
+    There are three cases:
+    1. full tuning without ds_zero3: state_dict = {"model.layers.*": ..., "v_head.summary.*": ...}
+    2. lora tuning without ds_zero3: state_dict = {"v_head.summary.*": ...}
+    3. under deepspeed zero3: state_dict = {"pretrained_model.model.layers.*": ..., "v_head.summary.*": ...}
+
+    We assume `stage3_gather_16bit_weights_on_model_save=true`.
+    """
+    if not isinstance(model.pretrained_model, (PreTrainedModel, PeftModel)):
+        return
+
+    if safe_serialization:
+        path_to_checkpoint = os.path.join(output_dir, SAFE_WEIGHTS_NAME)
+        with safe_open(path_to_checkpoint, framework="pt", device="cpu") as f:
+            state_dict: Dict[str, torch.Tensor] = {key: f.get_tensor(key) for key in f.keys()}
+    else:
+        path_to_checkpoint = os.path.join(output_dir, WEIGHTS_NAME)
+        state_dict: Dict[str, torch.Tensor] = torch.load(path_to_checkpoint, map_location="cpu")
+
+    os.remove(path_to_checkpoint)
+    decoder_state_dict, v_head_state_dict = {}, {}
+    for name, param in state_dict.items():
+        if name.startswith("v_head."):
+            v_head_state_dict[name] = param
+        else:
+            decoder_state_dict[name.replace("pretrained_model.", "", 1)] = param
+
+    model.pretrained_model.save_pretrained(
+        output_dir, state_dict=decoder_state_dict or None, safe_serialization=safe_serialization
+    )
+
+    if safe_serialization:
+        save_file(v_head_state_dict, os.path.join(output_dir, V_HEAD_SAFE_WEIGHTS_NAME), metadata={"format": "pt"})
+    else:
+        torch.save(v_head_state_dict, os.path.join(output_dir, V_HEAD_WEIGHTS_NAME))
+
+    logger.info("Value head model saved at: {}".format(output_dir))
+
+
+class FixValueHeadModelCallback(TrainerCallback):
+    r"""
+    A callback for fixing the checkpoint for valuehead models.
+    """
+
+    @override
+    def on_save(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called after a checkpoint save.
+        """
+        if args.should_save:
+            output_dir = os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step))
+            fix_valuehead_checkpoint(
+                model=kwargs.pop("model"), output_dir=output_dir, safe_serialization=args.save_safetensors
+            )
+
+
+class SaveProcessorCallback(TrainerCallback):
+    r"""
+    A callback for saving the processor.
+    """
+
+    def __init__(self, processor: "ProcessorMixin") -> None:
+        self.processor = processor
+
+    @override
+    def on_save(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if args.should_save:
+            output_dir = os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step))
+            getattr(self.processor, "image_processor").save_pretrained(output_dir)
+
+    @override
+    def on_train_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if args.should_save:
+            getattr(self.processor, "image_processor").save_pretrained(args.output_dir)
+
+
+class PissaConvertCallback(TrainerCallback):
+    r"""
+    A callback for converting the PiSSA adapter to a normal one.
+    """
+
+    @override
+    def on_train_begin(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        r"""
+        Event called at the beginning of training.
+        """
+        if args.should_save:
+            model = kwargs.pop("model")
+            pissa_init_dir = os.path.join(args.output_dir, "pissa_init")
+            logger.info("Initial PiSSA adapter will be saved at: {}.".format(pissa_init_dir))
+            if isinstance(model, PeftModel):
+                init_lora_weights = getattr(model.peft_config["default"], "init_lora_weights")
+                setattr(model.peft_config["default"], "init_lora_weights", True)
+                model.save_pretrained(pissa_init_dir, safe_serialization=args.save_safetensors)
+                setattr(model.peft_config["default"], "init_lora_weights", init_lora_weights)
+
+    @override
+    def on_train_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if args.should_save:
+            model = kwargs.pop("model")
+            pissa_init_dir = os.path.join(args.output_dir, "pissa_init")
+            pissa_backup_dir = os.path.join(args.output_dir, "pissa_backup")
+            pissa_convert_dir = os.path.join(args.output_dir, "pissa_converted")
+            logger.info("Converted PiSSA adapter will be saved at: {}.".format(pissa_convert_dir))
+            # 1. save a pissa backup with init_lora_weights: True
+            # 2. save a converted lora with init_lora_weights: pissa
+            # 3. load the pissa backup with init_lora_weights: True
+            # 4. delete the initial adapter and change init_lora_weights to pissa
+            if isinstance(model, PeftModel):
+                init_lora_weights = getattr(model.peft_config["default"], "init_lora_weights")
+                setattr(model.peft_config["default"], "init_lora_weights", True)
+                model.save_pretrained(pissa_backup_dir, safe_serialization=args.save_safetensors)
+                setattr(model.peft_config["default"], "init_lora_weights", init_lora_weights)
+                model.save_pretrained(
+                    pissa_convert_dir, safe_serialization=args.save_safetensors, convert_pissa_to_lora=pissa_init_dir
+                )  # TODO: use `path_initial_model_for_weight_conversion` (peft>=0.12.0)
+                model.load_adapter(pissa_backup_dir, "default", is_trainable=True)
+                model.set_adapter("default")
+                if "pissa_init" in model.peft_config.keys():  # backward compatibility (peft<0.12.0)
+                    model.delete_adapter("pissa_init")
+
+                setattr(model.peft_config["default"], "init_lora_weights", init_lora_weights)
+
+
+class LogCallback(TrainerCallback):
+    r"""
+    A callback for logging training and evaluation status.
+    """
+
+    def __init__(self) -> None:
+        # Progress
+        self.start_time = 0
+        self.cur_steps = 0
+        self.max_steps = 0
+        self.elapsed_time = ""
+        self.remaining_time = ""
+        self.thread_pool: Optional["ThreadPoolExecutor"] = None
+        # Status
+        self.aborted = False
+        self.do_train = False
+        # Web UI
+        self.webui_mode = os.environ.get("LLAMABOARD_ENABLED", "0").lower() in ["true", "1"]
+        if self.webui_mode:
+            signal.signal(signal.SIGABRT, self._set_abort)
+            self.logger_handler = LoggerHandler(os.environ.get("LLAMABOARD_WORKDIR"))
+            logging.root.addHandler(self.logger_handler)
+            transformers.logging.add_handler(self.logger_handler)
+
+    def _set_abort(self, signum, frame) -> None:
+        self.aborted = True
+
+    def _reset(self, max_steps: int = 0) -> None:
+        self.start_time = time.time()
+        self.cur_steps = 0
+        self.max_steps = max_steps
+        self.elapsed_time = ""
+        self.remaining_time = ""
+
+    def _timing(self, cur_steps: int) -> None:
+        cur_time = time.time()
+        elapsed_time = cur_time - self.start_time
+        avg_time_per_step = elapsed_time / cur_steps if cur_steps != 0 else 0
+        remaining_time = (self.max_steps - cur_steps) * avg_time_per_step
+        self.cur_steps = cur_steps
+        self.elapsed_time = str(timedelta(seconds=int(elapsed_time)))
+        self.remaining_time = str(timedelta(seconds=int(remaining_time)))
+
+    def _write_log(self, output_dir: str, logs: Dict[str, Any]) -> None:
+        with open(os.path.join(output_dir, TRAINER_LOG), "a", encoding="utf-8") as f:
+            f.write(json.dumps(logs) + "\n")
+
+    def _create_thread_pool(self, output_dir: str) -> None:
+        os.makedirs(output_dir, exist_ok=True)
+        self.thread_pool = ThreadPoolExecutor(max_workers=1)
+
+    def _close_thread_pool(self) -> None:
+        if self.thread_pool is not None:
+            self.thread_pool.shutdown(wait=True)
+            self.thread_pool = None
+
+    @override
+    def on_init_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if (
+            args.should_save
+            and os.path.exists(os.path.join(args.output_dir, TRAINER_LOG))
+            and args.overwrite_output_dir
+        ):
+            logger.warning("Previous trainer log in this folder will be deleted.")
+            os.remove(os.path.join(args.output_dir, TRAINER_LOG))
+
+    @override
+    def on_train_begin(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if args.should_save:
+            self.do_train = True
+            self._reset(max_steps=state.max_steps)
+            self._create_thread_pool(output_dir=args.output_dir)
+
+    @override
+    def on_train_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        self._close_thread_pool()
+
+    @override
+    def on_substep_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if self.aborted:
+            control.should_epoch_stop = True
+            control.should_training_stop = True
+
+    @override
+    def on_step_end(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if self.aborted:
+            control.should_epoch_stop = True
+            control.should_training_stop = True
+
+    @override
+    def on_evaluate(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if not self.do_train:
+            self._close_thread_pool()
+
+    @override
+    def on_predict(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if not self.do_train:
+            self._close_thread_pool()
+
+    @override
+    def on_log(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
+        if not args.should_save:
+            return
+
+        self._timing(cur_steps=state.global_step)
+        logs = dict(
+            current_steps=self.cur_steps,
+            total_steps=self.max_steps,
+            loss=state.log_history[-1].get("loss", None),
+            eval_loss=state.log_history[-1].get("eval_loss", None),
+            predict_loss=state.log_history[-1].get("predict_loss", None),
+            reward=state.log_history[-1].get("reward", None),
+            accuracy=state.log_history[-1].get("rewards/accuracies", None),
+            learning_rate=state.log_history[-1].get("learning_rate", None),
+            epoch=state.log_history[-1].get("epoch", None),
+            percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,
+            elapsed_time=self.elapsed_time,
+            remaining_time=self.remaining_time,
+        )
+        if state.num_input_tokens_seen:
+            logs["throughput"] = round(state.num_input_tokens_seen / (time.time() - self.start_time), 2)
+            logs["total_tokens"] = state.num_input_tokens_seen
+
+        if os.environ.get("RECORD_VRAM", "0").lower() in ["true", "1"]:
+            vram_allocated, vram_reserved = get_peak_memory()
+            logs["vram_allocated"] = round(vram_allocated / 1024 / 1024 / 1024, 2)
+            logs["vram_reserved"] = round(vram_reserved / 1024 / 1024 / 1024, 2)
+
+        logs = {k: v for k, v in logs.items() if v is not None}
+        if self.webui_mode and all(key in logs for key in ["loss", "learning_rate", "epoch"]):
+            logger.info(
+                "{{'loss': {:.4f}, 'learning_rate': {:2.4e}, 'epoch': {:.2f}, 'throughput': {}}}".format(
+                    logs["loss"], logs["learning_rate"], logs["epoch"], logs.get("throughput", "N/A")
+                )
+            )
+
+        if self.thread_pool is not None:
+            self.thread_pool.submit(self._write_log, args.output_dir, logs)
+
+    @override
+    def on_prediction_step(
+        self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs
+    ):
+        if self.do_train:
+            return
+
+        if self.aborted:
+            sys.exit(0)
+
+        if not args.should_save:
+            return
+
+        eval_dataloader = kwargs.pop("eval_dataloader", None)
+        if has_length(eval_dataloader):
+            if self.max_steps == 0:
+                self._reset(max_steps=len(eval_dataloader))
+                self._create_thread_pool(output_dir=args.output_dir)
+
+            self._timing(cur_steps=self.cur_steps + 1)
+            if self.cur_steps % 5 == 0 and self.thread_pool is not None:
+                logs = dict(
+                    current_steps=self.cur_steps,
+                    total_steps=self.max_steps,
+                    percentage=round(self.cur_steps / self.max_steps * 100, 2) if self.max_steps != 0 else 100,
+                    elapsed_time=self.elapsed_time,
+                    remaining_time=self.remaining_time,
+                )
+                self.thread_pool.submit(self._write_log, args.output_dir, logs)

src/llamafactory/train/dpo/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 .workflow import run_dpo
+
+
+__all__ = ["run_dpo"]

src/llamafactory/train/dpo/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/trl/trainer/dpo_trainer.py
+#
+# 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 warnings
+from collections import defaultdict
+from contextlib import nullcontext
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import Trainer
+from trl import DPOTrainer
+from trl.trainer import disable_dropout_in_model
+from typing_extensions import override
+
+from ...extras.constants import IGNORE_INDEX
+from ..callbacks import PissaConvertCallback, SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, ProcessorMixin
+
+    from ...hparams import FinetuningArguments
+
+
+class CustomDPOTrainer(DPOTrainer):
+    def __init__(
+        self,
+        model: Union["PreTrainedModel", torch.nn.Module],
+        ref_model: Optional[Union["PreTrainedModel", torch.nn.Module]],
+        finetuning_args: "FinetuningArguments",
+        processor: Optional["ProcessorMixin"],
+        disable_dropout: bool = True,
+        **kwargs,
+    ):
+        if disable_dropout:
+            disable_dropout_in_model(model)
+            if ref_model is not None:
+                disable_dropout_in_model(ref_model)
+
+        self.finetuning_args = finetuning_args
+        self.f_divergence_type = "reverse_kl"
+        self.reference_free = False
+        self.use_dpo_data_collator = True  # hack to avoid warning
+        self.generate_during_eval = False  # disable at evaluation
+        self.label_pad_token_id = IGNORE_INDEX
+        self.padding_value = 0
+        self.is_encoder_decoder = model.config.is_encoder_decoder
+        self.precompute_ref_log_probs = False
+        self._precomputed_train_ref_log_probs = False
+        self._precomputed_eval_ref_log_probs = False
+        self._peft_has_been_casted_to_bf16 = False
+
+        self.ref_model = ref_model
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+
+        # dpo hyperparams
+        self.beta = finetuning_args.pref_beta
+        self.loss_type = finetuning_args.pref_loss
+        self.ftx_gamma = finetuning_args.pref_ftx
+        self.label_smoothing = finetuning_args.dpo_label_smoothing
+        self.simpo_gamma = finetuning_args.simpo_gamma
+
+        Trainer.__init__(self, model=model, **kwargs)
+        if not hasattr(self, "accelerator"):
+            raise AttributeError("Please update `transformers`.")
+
+        warnings.simplefilter("ignore")  # remove gc warnings on ref model
+
+        if ref_model is not None:
+            if self.is_deepspeed_enabled:
+                if not (
+                    getattr(ref_model, "is_loaded_in_8bit", False) or getattr(ref_model, "is_loaded_in_4bit", False)
+                ):  # quantized models are already set on the correct device
+                    self.ref_model = self._prepare_deepspeed(self.ref_model)
+            else:
+                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
+                self.ref_model.eval()
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.pissa_convert:
+            self.callback_handler.add_callback(PissaConvertCallback)
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
+    ) -> "torch.optim.lr_scheduler.LRScheduler":
+        create_custom_scheduler(self.args, num_training_steps, optimizer)
+        return super().create_scheduler(num_training_steps, optimizer)
+
+    def odds_ratio_loss(self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor") -> "torch.Tensor":
+        r"""
+        Computes ORPO's odds ratio (OR) loss for batched log probabilities of the policy model.
+        """
+        log_odds = (chosen_logps - rejected_logps) - (
+            torch.log1p(-torch.exp(chosen_logps)) - torch.log1p(-torch.exp(rejected_logps))
+        )
+        sft_loss = -chosen_logps
+        odds_ratio_loss = -F.logsigmoid(log_odds)
+        orpo_loss = sft_loss + self.beta * odds_ratio_loss
+        return orpo_loss
+
+    def simpo_loss(self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor") -> "torch.Tensor":
+        r"""
+        Computes SimPO loss for batched log probabilities of the policy model.
+        """
+        pi_logratios = chosen_logps - rejected_logps
+        gamma_logratios = self.simpo_gamma / self.beta
+        logits = pi_logratios - gamma_logratios
+        simpo_loss = -F.logsigmoid(self.beta * logits)
+        return simpo_loss
+
+    def compute_preference_loss(
+        self,
+        policy_chosen_logps: "torch.Tensor",
+        policy_rejected_logps: "torch.Tensor",
+        reference_chosen_logps: Optional["torch.Tensor"],
+        reference_rejected_logps: Optional["torch.Tensor"],
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes loss for preference learning.
+        """
+        if not self.finetuning_args.use_ref_model:
+            if self.loss_type == "orpo":
+                losses = self.odds_ratio_loss(policy_chosen_logps, policy_rejected_logps)
+            elif self.loss_type == "simpo":
+                losses = self.simpo_loss(policy_chosen_logps, policy_rejected_logps)
+            else:
+                raise NotImplementedError("Unknown loss type: {}.".format(self.loss_type))
+
+            chosen_rewards = self.beta * policy_chosen_logps.to(self.accelerator.device).detach()
+            rejected_rewards = self.beta * policy_rejected_logps.to(self.accelerator.device).detach()
+        else:
+            losses, chosen_rewards, rejected_rewards = self.dpo_loss(
+                policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps
+            )
+
+        return losses, chosen_rewards, rejected_rewards
+
+    @override
+    def concatenated_forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes the sum log probabilities of the labels under given logits if loss_type is not IPO, ORPO or SimPO.
+
+        Otherwise the average log probabilities.
+        """
+        if self.finetuning_args.use_ref_model:
+            batch = {k: v.detach().clone() for k, v in batch.items()}  # avoid error
+
+        all_logits: "torch.Tensor" = model(**batch, return_dict=True, use_cache=False).logits.to(torch.float32)
+        all_logps, valid_length = get_batch_logps(logits=all_logits, labels=batch["labels"])
+        if self.loss_type in ["ipo", "orpo", "simpo"]:
+            all_logps = all_logps / valid_length
+
+        batch_size = batch["input_ids"].size(0) // 2
+        chosen_logps, rejected_logps = all_logps.split(batch_size, dim=0)
+        chosen_logits, rejected_logits = all_logits.split(batch_size, dim=0)
+        chosen_length, _ = valid_length.split(batch_size, dim=0)
+        return chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_logps / chosen_length
+
+    @override
+    def compute_reference_log_probs(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple[Optional["torch.Tensor"], Optional["torch.Tensor"]]:
+        r"""
+        Computes log probabilities of the reference model.
+        """
+        if not self.finetuning_args.use_ref_model:
+            return None, None
+
+        if self.ref_model is None:
+            ref_model = model
+            ref_context = self.accelerator.unwrap_model(model).disable_adapter()
+        else:
+            ref_model = self.ref_model
+            ref_context = nullcontext()
+
+        with torch.no_grad(), ref_context:
+            reference_chosen_logps, reference_rejected_logps, *_ = self.concatenated_forward(ref_model, batch)
+
+        return reference_chosen_logps, reference_rejected_logps
+
+    @override
+    def get_batch_loss_metrics(
+        self,
+        model: "PreTrainedModel",
+        batch: Dict[str, "torch.Tensor"],
+        train_eval: Literal["train", "eval"] = "train",
+    ) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
+        r"""
+        Computes the DPO loss and other metrics for the given batch of inputs for train or test.
+        """
+        metrics = {}
+        (
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits,
+            policy_rejected_logits,
+            policy_chosen_logps_avg,
+        ) = self.concatenated_forward(model, batch)
+
+        reference_chosen_logps, reference_rejected_logps = self.compute_reference_log_probs(model, batch)
+        losses, chosen_rewards, rejected_rewards = self.compute_preference_loss(
+            policy_chosen_logps,
+            policy_rejected_logps,
+            reference_chosen_logps,
+            reference_rejected_logps,
+        )
+        sft_loss = -policy_chosen_logps_avg
+        if self.ftx_gamma > 1e-6:
+            losses += self.ftx_gamma * sft_loss
+
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.mean().cpu()
+        metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.mean().cpu()
+        metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.mean().cpu()
+        metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).mean().cpu()
+        metrics["{}logps/rejected".format(prefix)] = policy_rejected_logps.detach().mean().cpu()
+        metrics["{}logps/chosen".format(prefix)] = policy_chosen_logps.detach().mean().cpu()
+        metrics["{}logits/rejected".format(prefix)] = policy_rejected_logits.detach().mean().cpu()
+        metrics["{}logits/chosen".format(prefix)] = policy_chosen_logits.detach().mean().cpu()
+        if self.loss_type == "orpo":
+            metrics["{}sft_loss".format(prefix)] = sft_loss.detach().mean().cpu()
+            metrics["{}odds_ratio_loss".format(prefix)] = ((losses - sft_loss) / self.beta).detach().mean().cpu()
+
+        return losses.mean(), metrics

src/llamafactory/train/dpo/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/examples/scripts/dpo.py
+#
+# 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 typing import TYPE_CHECKING, List, Optional
+
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, get_template_and_fix_tokenizer
+from ...extras.constants import IGNORE_INDEX
+from ...extras.ploting import plot_loss
+from ...hparams import ModelArguments
+from ...model import load_model, load_tokenizer
+from ..trainer_utils import create_modelcard_and_push, create_ref_model
+from .trainer import CustomDPOTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments
+
+
+def run_dpo(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="rm", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+
+    data_collator = PairwiseDataCollatorWithPadding(
+        template=template,
+        pad_to_multiple_of=8,
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
+        **tokenizer_module,
+    )
+
+    # Create reference model
+    if finetuning_args.use_ref_model:
+        if finetuning_args.ref_model is None and (not training_args.do_train):  # use the model itself
+            ref_model = model
+        else:
+            ref_model = create_ref_model(model_args, finetuning_args)
+    else:
+        ref_model = None
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
+
+    # Initialize our Trainer
+    trainer = CustomDPOTrainer(
+        model=model,
+        ref_model=ref_model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **dataset_module,
+        **tokenizer_module,
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "rewards/accuracies"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        if id(model) == id(ref_model):  # unable to compute rewards if reference model is the model itself
+            remove_keys = [key for key in metrics.keys() if "rewards" in key]
+            for key in remove_keys:
+                metrics.pop(key)
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Create model card
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

src/llamafactory/train/kto/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# 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 .workflow import run_kto
+
+
+__all__ = ["run_kto"]

src/llamafactory/train/kto/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/trl/trainer/kto_trainer.py
+#
+# 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 warnings
+from collections import defaultdict
+from contextlib import nullcontext
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
+
+import torch
+from transformers import Trainer
+from trl import KTOTrainer
+from trl.trainer import disable_dropout_in_model
+from typing_extensions import override
+
+from ...extras.constants import IGNORE_INDEX
+from ..callbacks import SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps
+
+
+if TYPE_CHECKING:
+    import torch.utils.data
+    from transformers import PreTrainedModel, ProcessorMixin
+
+    from ...hparams import FinetuningArguments
+
+
+class CustomKTOTrainer(KTOTrainer):
+    def __init__(
+        self,
+        model: Union["PreTrainedModel", torch.nn.Module],
+        ref_model: Optional[Union["PreTrainedModel", torch.nn.Module]],
+        finetuning_args: "FinetuningArguments",
+        processor: Optional["ProcessorMixin"],
+        disable_dropout: bool = True,
+        **kwargs,
+    ):
+        if disable_dropout:
+            disable_dropout_in_model(model)
+            if ref_model is not None:
+                disable_dropout_in_model(ref_model)
+
+        self.finetuning_args = finetuning_args
+        self.reference_free = False
+        self.use_dpo_data_collator = True  # hack to avoid warning
+        self.generate_during_eval = False  # disable at evaluation
+        self.label_pad_token_id = IGNORE_INDEX
+        self.padding_value = 0
+        self.is_encoder_decoder = model.config.is_encoder_decoder
+        self.precompute_ref_log_probs = False
+        self._precomputed_train_ref_log_probs = False
+        self._precomputed_eval_ref_log_probs = False
+        self._peft_has_been_casted_to_bf16 = False
+
+        self.ref_model = ref_model
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+
+        # kto hyperparams
+        self.beta = finetuning_args.pref_beta
+        self.desirable_weight = finetuning_args.kto_chosen_weight
+        self.undesirable_weight = finetuning_args.kto_rejected_weight
+        self.ftx_gamma = finetuning_args.pref_ftx
+
+        Trainer.__init__(self, model=model, **kwargs)
+        if not hasattr(self, "accelerator"):
+            raise AttributeError("Please update `transformers`.")
+
+        warnings.simplefilter("ignore")  # remove gc warnings on ref model
+
+        if ref_model is not None:
+            if self.is_deepspeed_enabled:
+                if not (
+                    getattr(ref_model, "is_loaded_in_8bit", False) or getattr(ref_model, "is_loaded_in_4bit", False)
+                ):  # quantized models are already set on the correct device
+                    self.ref_model = self._prepare_deepspeed(self.ref_model)
+            else:
+                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
+                self.ref_model.eval()
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
+    ) -> "torch.optim.lr_scheduler.LRScheduler":
+        create_custom_scheduler(self.args, num_training_steps, optimizer)
+        return super().create_scheduler(num_training_steps, optimizer)
+
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        r"""
+        Replaces the sequential sampler of KTO Trainer created by trl with the random sampler.
+        """
+        return Trainer._get_train_sampler(self)
+
+    @override
+    def forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"], prefix: Literal["", "kl_"] = ""
+    ) -> Tuple["torch.Tensor", "torch.Tensor"]:
+        r"""
+        Runs forward pass and computes the log probabilities.
+        """
+        batch = {k: v.detach().clone() for k, v in batch.items()}  # avoid error
+        model_inputs = {
+            "input_ids": batch["{}input_ids".format(prefix)],
+            "attention_mask": batch["{}attention_mask".format(prefix)],
+        }
+        if "{}token_type_ids".format(prefix) in batch:
+            model_inputs["token_type_ids"] = batch["{}token_type_ids".format(prefix)]
+
+        if "pixel_values" in batch:
+            model_inputs["pixel_values"] = batch["pixel_values"]
+
+        if "image_grid_thw" in batch:
+            model_inputs["image_grid_thw"] = batch["image_grid_thw"]
+
+        logits = model(**model_inputs, return_dict=True, use_cache=False).logits.to(torch.float32)
+        logps, valid_length = get_batch_logps(logits=logits, labels=batch["{}labels".format(prefix)])
+        return logps, logps / valid_length
+
+    @override
+    def concatenated_forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        target_logps, target_logps_avg = self.forward(model, batch)
+        with torch.no_grad():
+            kl_logps, _ = self.forward(model, batch, prefix="kl_")
+
+        if len(target_logps) != len(batch["kto_tags"]):
+            raise ValueError("Mismatched shape of inputs and labels.")
+
+        chosen_logps = target_logps[batch["kto_tags"]]
+        rejected_logps = target_logps[~batch["kto_tags"]]
+        chosen_logps_avg = target_logps_avg[batch["kto_tags"]]
+        return chosen_logps, rejected_logps, kl_logps, chosen_logps_avg
+
+    @override
+    def compute_reference_log_probs(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes log probabilities of the reference model.
+        """
+        if self.ref_model is None:
+            ref_model = model
+            ref_context = self.accelerator.unwrap_model(model).disable_adapter()
+        else:
+            ref_model = self.ref_model
+            ref_context = nullcontext()
+
+        with torch.no_grad(), ref_context:
+            reference_chosen_logps, reference_rejected_logps, reference_kl_logps, _ = self.concatenated_forward(
+                ref_model, batch
+            )
+
+        return reference_chosen_logps, reference_rejected_logps, reference_kl_logps
+
+    @override
+    def get_batch_loss_metrics(
+        self,
+        model: "PreTrainedModel",
+        batch: Dict[str, "torch.Tensor"],
+    ) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
+        r"""
+        Computes the DPO loss and other metrics for the given batch of inputs for train or test.
+        """
+        metrics = {}
+        policy_chosen_logps, policy_rejected_logps, policy_kl_logps, policy_chosen_logps_avg = (
+            self.concatenated_forward(model, batch)
+        )
+        reference_chosen_logps, reference_rejected_logps, reference_kl_logps = self.compute_reference_log_probs(
+            model, batch
+        )
+        losses, chosen_rewards, rejected_rewards, kl = self.kto_loss(
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_kl_logps,
+            reference_chosen_logps,
+            reference_rejected_logps,
+            reference_kl_logps,
+        )
+        losses = losses.nanmean()
+
+        if self.ftx_gamma > 1e-6 and len(policy_chosen_logps) > 0:  # remember to rescale
+            sft_loss = -policy_chosen_logps_avg
+            losses += self.ftx_gamma * sft_loss.nanmean() / len(policy_chosen_logps) * len(batch["labels"])
+
+        num_chosen = torch.Tensor([len(chosen_rewards)]).to(self.accelerator.device)
+        num_rejected = torch.Tensor([len(rejected_rewards)]).to(self.accelerator.device)
+
+        all_num_chosen = self.accelerator.gather(num_chosen).sum().item()
+        all_num_rejected = self.accelerator.gather(num_rejected).sum().item()
+
+        if all_num_chosen > 0:
+            metrics["rewards/chosen_sum"] = self.accelerator.gather(chosen_rewards.nansum()).nansum().item()
+            metrics["logps/chosen_sum"] = self.accelerator.gather(policy_chosen_logps.nansum()).nansum().item()
+            metrics["count/chosen"] = all_num_chosen
+
+        if all_num_rejected > 0:
+            metrics["rewards/rejected_sum"] = self.accelerator.gather(rejected_rewards.nansum()).nansum().item()
+            metrics["logps/rejected_sum"] = self.accelerator.gather(policy_rejected_logps.nansum()).nansum().item()
+            metrics["count/rejected"] = all_num_rejected
+
+        metrics["kl"] = kl.item()
+
+        return losses, metrics