Skip to content

GitLab

Unverified Commit fb2f74e2 authored by Nicolas Patry's avatar Nicolas Patry Committed by GitHub
Browse files

Refactor dead code - Removing all `flash_xxx.py` files. (#2166) 

* Refactor dead code.

* First working step.

* Remove a lot of duplicated code.

* More dead code.

* More cleanup.

* Fix Santacoder test.

* Fixing the simple tests.

* Fixing sharding.

* Fixes for VLM.

* Fixing santacoder (num_kv_heads hardcoded).

* Removing more dead code.

* Fixing `config.n_head`.

* Stopping earlier because of `<end_of_utterance>` in idefics2.

* Addresses comments.

* Removing the dead code.

* Fuse back mistral into FlashCausalLM.

* Finish removal.

* Fixing docs + causal_lm `batch_class`.

* Fixing docs + causal.lm.

* Add default to Gemma Causality.

* Default value for gemma/gemma2.

* Wrong default.
parent c6bcadf8
Hide whitespace changes
Inline Side-by-side
Showing with 2 additions and 1697 deletions
server/text_generation_server/models/flash_gpt2.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoConfig, AutoTokenizer, GenerationConfig
from transformers.models.gpt2 import GPT2Tokenizer
from typing import Optional
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_gpt2_modeling import (
FlashGPT2ForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
class FlashGPT2(FlashCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashGPT2 is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["gptq", "awq", "marlin"]:
weights._set_gptq_params(model_id, revision)
prefix = ""
model = FlashGPT2ForCausalLM(prefix, config, weights)
torch.distributed.barrier(group=self.process_group)
super(FlashGPT2, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
num_layers=len(model.model.layers),
num_kv_heads=model.model.num_heads,
head_size=model.model.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
server/text_generation_server/models/flash_llama.py deleted 100644 → 0
View file @ c6bcadf8
import os
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoConfig, AutoTokenizer, GenerationConfig
from typing import Optional, Tuple, Dict, List
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_llama_modeling import (
FlashLlamaForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
hub,
)
tracer = trace.get_tracer(__name__)
from text_generation_server.utils.import_utils import SYSTEM
ADAPTER_LAYERS = [
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
]
ROW_PARALLEL = {"o_proj", "down_proj", "lm_head"}
class FlashLlama(FlashCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
lora_adapter_ids: Optional[list] = [],
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashLlama is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
try:
generation_config = GenerationConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
if isinstance(generation_config.eos_token_id, (list, set)):
# TODO Huge hack
tokenizer._eos_token_ids = set(generation_config.eos_token_id)
except Exception:
pass
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["awq", "exl2", "gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
prefix = ""
model = FlashLlamaForCausalLM(prefix, config, weights)
torch.distributed.barrier(group=self.process_group)
super(FlashLlama, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
num_layers=len(model.model.layers),
num_kv_heads=model.model.num_key_value_heads,
head_size=model.model.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
@property
def supports_adapter_loading(self) -> bool:
return True
def adapter_target_to_layer(self) -> Dict[str, Tuple[str, torch.Tensor]]:
layer_weights = {}
prefix = "model.layers"
# This accounts for VLMs (e.g. LlavaNext, Idefics2)
# that have a language_model inside of the larger model.
if hasattr(self.model, "language_model"):
_model = self.model.language_model
elif hasattr(self.model, "text_model"):
_model = self.model.text_model
else:
_model = self.model
for i, layer in enumerate(_model.model.layers):
layer_weights[(i, "q_proj")] = (
f"{prefix}.{i}.self_attn.q_proj",
layer.self_attn.query_key_value,
)
layer_weights[(i, "k_proj")] = (
f"{prefix}.{i}.self_attn.k_proj",
layer.self_attn.query_key_value,
)
layer_weights[(i, "v_proj")] = (
f"{prefix}.{i}.self_attn.v_proj",
layer.self_attn.query_key_value,
)
layer_weights[(i, "o_proj")] = (
f"{prefix}.{i}.self_attn.o_proj",
layer.self_attn.o_proj,
)
layer_weights[(i, "gate_proj")] = (
f"{prefix}.{i}.mlp.gate_proj",
layer.mlp.gate_up_proj,
)
layer_weights[(i, "up_proj")] = (
f"{prefix}.{i}.mlp.up_proj",
layer.mlp.gate_up_proj,
)
layer_weights[(i, "down_proj")] = (
f"{prefix}.{i}.mlp.down_proj",
layer.mlp.down_proj,
)
layer_weights[(0, "lm_head")] = ("lm_head", _model.lm_head)
return layer_weights
@property
def adapter_layers(self) -> List[str]:
return ADAPTER_LAYERS
@property
def default_traced_adapter_layers(self) -> List[str]:
return ["q_proj", "v_proj"]
def get_num_layers_for_type(self, layer_type: str) -> int:
return 1 if layer_type == "lm_head" else len(self.model.model.layers)
def is_row_parallel(self, layer_type: str) -> bool:
return layer_type in ROW_PARALLEL
server/text_generation_server/models/flash_mistral.py
View file @ fb2f74e2
import torch import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoTokenizer, AutoConfig
from typing import Optional, Tuple, Dict, List from typing import Optional, Tuple, Dict, List
from text_generation_server.models import FlashCausalLM from text_generation_server.models import FlashCausalLM
from text_generation_server.models.flash_causal_lm import set_sliding_window
from text_generation_server.models.custom_modeling.flash_mistral_modeling import (
FlashMistralForCausalLM,
MistralConfig,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
ADAPTER_LAYERS = [ ADAPTER_LAYERS = [
...@@ -33,88 +16,7 @@ ADAPTER_LAYERS = [ ...@@ -33,88 +16,7 @@ ADAPTER_LAYERS = [
ROW_PARALLEL = {"o_proj", "down_proj", "lm_head"} ROW_PARALLEL = {"o_proj", "down_proj", "lm_head"}
class BaseFlashMistral(FlashCausalLM): class FlashMistral(FlashCausalLM):
def __init__(
self,
model_cls,
model_id: str,
config_cls=AutoConfig,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
tokenizer_class=AutoTokenizer,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashMistral is only available on GPU")
tokenizer = tokenizer_class.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = config_cls.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
# Set context windows
if getattr(config, "sliding_window", None) is not None:
set_sliding_window(config.sliding_window)
else:
config.sliding_window = None
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["gptq", "awq", "marlin"]:
weights._set_gptq_params(model_id, revision)
prefix = ""
model = model_cls(prefix, config, weights)
self.cuda_graphs = {}
torch.distributed.barrier(group=self.process_group)
num_layers, num_kv_heads, head_size = self.get_layer_config(model)
super().__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
num_layers=num_layers,
num_kv_heads=num_kv_heads,
head_size=head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
sliding_window=config.sliding_window,
)
def get_layer_config(self, model) -> Tuple[int, int, int]:
return (
len(model.model.layers),
model.model.num_key_value_heads,
model.model.head_size,
)
@property @property
def supports_adapter_loading(self) -> bool: def supports_adapter_loading(self) -> bool:
return True return True
...@@ -126,9 +28,7 @@ class BaseFlashMistral(FlashCausalLM): ...@@ -126,9 +28,7 @@ class BaseFlashMistral(FlashCausalLM):
# This accounts for VLMs (e.g. LlavaNext, Idefics2) # This accounts for VLMs (e.g. LlavaNext, Idefics2)
# that have a language_model inside of the larger model. # that have a language_model inside of the larger model.
if hasattr(self.model, "language_model"): if hasattr(self.model, "text_model"):
_model = self.model.language_model
elif hasattr(self.model, "text_model"):
_model = self.model.text_model _model = self.model.text_model
else: else:
_model = self.model _model = self.model
...@@ -183,25 +83,3 @@ class BaseFlashMistral(FlashCausalLM): ...@@ -183,25 +83,3 @@ class BaseFlashMistral(FlashCausalLM):
def is_row_parallel(self, layer_type: str) -> bool: def is_row_parallel(self, layer_type: str) -> bool:
return layer_type in ROW_PARALLEL return layer_type in ROW_PARALLEL
class FlashMistral(BaseFlashMistral):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
super(FlashMistral, self).__init__(
config_cls=MistralConfig,
model_cls=FlashMistralForCausalLM,
model_id=model_id,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
server/text_generation_server/models/flash_mixtral.py deleted 100644 → 0
View file @ c6bcadf8
import torch
from typing import Optional
from text_generation_server.models.flash_mistral import BaseFlashMistral
from text_generation_server.models.custom_modeling.flash_mixtral_modeling import (
MixtralConfig,
FlashMixtralForCausalLM,
)
class FlashMixtral(BaseFlashMistral):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
super(FlashMixtral, self).__init__(
config_cls=MixtralConfig,
model_cls=FlashMixtralForCausalLM,
model_id=model_id,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
server/text_generation_server/models/flash_neox.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoTokenizer, AutoConfig
from typing import Optional
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_neox_modeling import (
FlashGPTNeoXForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
class FlashNeoXSharded(FlashCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashNeoX is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames, device=device, dtype=dtype, process_group=self.process_group
)
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = FlashGPTNeoXForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(FlashNeoXSharded, self).__init__(
model_id=model_id,
model=model.to(device),
tokenizer=tokenizer,
num_layers=len(model.gpt_neox.layers),
num_kv_heads=model.gpt_neox.num_heads,
head_size=model.gpt_neox.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
server/text_generation_server/models/flash_phi.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoConfig, AutoTokenizer
from typing import Optional
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_phi_modeling import (
FlashPhiForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
class FlashPhi(FlashCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashPhi is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["gptq", "awq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = FlashPhiForCausalLM(config, weights)
if speculator:
from text_generation_server.utils.medusa import MedusaModel
from huggingface_hub import hf_hub_download
import json
import os
from pathlib import Path
is_local_model = (
Path(speculator).exists() and Path(speculator).is_dir()
) or os.getenv("WEIGHTS_CACHE_OVERRIDE", None) is not None
if not is_local_model:
medusa_config = hf_hub_download(
speculator, revision=revision, filename="config.json"
)
medusa_head = hf_hub_download(
speculator, revision=revision, filename="medusa_lm_head.pt"
)
else:
medusa_config = str(Path(speculator) / "config.json")
medusa_head = str(Path(speculator) / "medusa_lm_head.pt")
with open(medusa_config, "r") as f:
config = json.load(f)
medusa_sf = medusa_head[: -len(".pt")] + ".safetensors"
weights = Weights(
[medusa_sf], device, dtype, process_group=self.process_group
)
lm_head = model.lm_head
model.lm_head = MedusaModel(config, weights, lm_head)
torch.distributed.barrier(group=self.process_group)
super(FlashPhi, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
num_layers=len(model.model.layers),
num_kv_heads=model.model.num_key_value_heads,
head_size=model.model.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
server/text_generation_server/models/flash_qwen2.py deleted 100644 → 0
View file @ c6bcadf8
import math
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoTokenizer, AutoConfig
from typing import Optional
from text_generation_server.models.flash_mistral import (
BaseFlashMistral,
set_sliding_window,
)
from text_generation_server.models.custom_modeling.flash_qwen2_modeling import (
Qwen2ForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
class FlashQwen2(BaseFlashMistral):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashQwen2 is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
# Set context windows
if config.sliding_window is not None:
set_sliding_window(config.sliding_window)
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["gptq", "awq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = Qwen2ForCausalLM(config, weights)
self.cuda_graphs = {}
torch.distributed.barrier(group=self.process_group)
super(BaseFlashMistral, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
num_layers=len(model.model.layers),
num_kv_heads=model.model.num_key_value_heads,
head_size=model.model.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
sliding_window=config.sliding_window,
)
server/text_generation_server/models/flash_rw.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoTokenizer
from typing import Optional
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_rw_modeling import (
RWConfig,
FlashRWForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
class FlashRWSharded(FlashCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashRW is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = RWConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames,
device,
dtype,
process_group=self.process_group,
aliases={
"lm_head.weight": ["transformer.word_embeddings.weight"],
"transformer.word_embeddings.weight": ["lm_head.weight"],
},
)
config.quantize = quantize
config.speculator = speculator
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = FlashRWForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(FlashRWSharded, self).__init__(
model_id=model_id,
model=model.to(device),
tokenizer=tokenizer,
num_layers=len(model.transformer.h),
num_kv_heads=model.transformer.cache_size,
head_size=model.transformer.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
server/text_generation_server/models/flash_santacoder.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from opentelemetry import trace
from transformers import AutoTokenizer, AutoConfig
from typing import Optional, List
import json
import os
from huggingface_hub import hf_hub_download
from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_santacoder_modeling import (
FlashSantacoderForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
from text_generation_server.utils.import_utils import SYSTEM
tracer = trace.get_tracer(__name__)
class FlashSantacoderSharded(FlashCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
elif SYSTEM == "ipex":
if hasattr(torch, "xpu") and torch.xpu.is_available():
device = torch.device(f"xpu:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.bfloat16 if dtype is None else dtype
else:
raise NotImplementedError("FlashSantacoderSharded is only available on GPU")
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id,
revision=revision,
trust_remote_code=True,
)
config.quantize = quantize
config.speculator = speculator
config.transpose = config.architectures[0].startswith("GPT2")
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames,
device=device,
dtype=dtype,
process_group=self.process_group,
aliases={"transformer.wte.weight": ["lm_head.weight"]},
)
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = FlashSantacoderForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(FlashSantacoderSharded, self).__init__(
model_id=model_id,
model=model.to(device),
tokenizer=tokenizer,
num_layers=len(model.transformer.h),
num_kv_heads=1,
head_size=model.transformer.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
def decode(self, generated_ids: List[int]) -> str:
# Do not skip special tokens as they are used for custom parsing rules of the generated text
return self.tokenizer.decode(
generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)
server/text_generation_server/models/flash_starcoder2.py deleted 100644 → 0
View file @ c6bcadf8
import math
import torch
from typing import Optional
from transformers.models.gpt2 import GPT2TokenizerFast
from text_generation_server.models.flash_mistral import (
BaseFlashMistral,
set_sliding_window,
)
from text_generation_server.models.custom_modeling.flash_starcoder2_modeling import (
Starcoder2Config,
FlashStarcoder2ForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
# Starcoder2 has the same base as Mistral
class FlashStarcoder2(BaseFlashMistral):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
raise NotImplementedError("FlashStarcoder2 is only available on GPU")
tokenizer = GPT2TokenizerFast.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = Starcoder2Config.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
config.quantize = quantize
config.speculator = speculator
# Set context windows
if config.sliding_window is not None:
set_sliding_window(config.sliding_window)
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["gptq", "awq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = FlashStarcoder2ForCausalLM(config, weights)
self.cuda_graphs = {}
torch.distributed.barrier(group=self.process_group)
super(BaseFlashMistral, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
num_layers=len(model.model.layers),
num_kv_heads=model.model.num_key_value_heads,
head_size=model.model.head_size,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
sliding_window=config.sliding_window,
)
server/text_generation_server/models/galactica.py
View file @ fb2f74e2
...@@ -162,83 +162,3 @@ class GalacticaCausalLMBatch(CausalLMBatch): ...@@ -162,83 +162,3 @@ class GalacticaCausalLMBatch(CausalLMBatch):
padding_right_offset=padding_right_offset, padding_right_offset=padding_right_offset,
max_tokens=max_tokens, max_tokens=max_tokens,
) )
class GalacticaSharded(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id,
revision=revision,
tp_parallel=True,
trust_remote_code=trust_remote_code,
)
config.quantize = quantize
tokenizer.pad_token_id = config.pad_token_id
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames, device=device, dtype=dtype, process_group=self.process_group
)
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = OPTForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
@property
def batch_type(self) -> Type[CausalLMBatch]:
return GalacticaCausalLMBatch
def decode(self, generated_ids: List[int]) -> str:
# Do not skip special tokens as they are used for custom parsing rules of the generated text
return self.tokenizer.decode(
generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
):
outputs, speculative_logits = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
use_cache=True,
)
return outputs.logits, speculative_logits, outputs.past_key_values
server/text_generation_server/models/gpt_neox.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from typing import Optional
from transformers import (
AutoTokenizer,
AutoConfig,
)
from text_generation_server.models import CausalLM
from text_generation_server.models.custom_modeling.neox_modeling import (
GPTNeoxForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
class GPTNeoxSharded(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
tokenizer.pad_token = tokenizer.eos_token
config = AutoConfig.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
)
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames, device=device, dtype=dtype, process_group=self.process_group
)
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = GPTNeoxForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
):
outputs, speculative_logits = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
use_cache=True,
)
return outputs.logits, speculative_logits, outputs.past_key_values
server/text_generation_server/models/idefics2.py deleted 100644 → 0
View file @ c6bcadf8
import torch
from typing import Optional, Tuple
from transformers import (
AutoProcessor,
)
from text_generation_server.models.custom_modeling.idefics2 import (
Idefics2ForConditionalGeneration,
)
from text_generation_server.models.vlm_causal_lm import VlmCausalLM
class Idefics2(VlmCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.processor = AutoProcessor.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
# XXX: Extremely important to cap resolution in order to limit
# VRAM usage.
size={"longest_edge": 448, "shortest_edge": 378},
)
super().__init__(
model_cls=Idefics2ForConditionalGeneration,
model_id=model_id,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
def get_layer_config(self, model) -> Tuple[int, int, int]:
return (
len(model.text_model.model.layers),
model.text_model.model.num_key_value_heads,
model.text_model.model.head_size,
)
def max_past(self) -> Optional[int]:
return getattr(self.model.text_model, "max_past", None)
server/text_generation_server/models/llava_next.py deleted 100644 → 0
View file @ c6bcadf8
import torch
from typing import Optional, Tuple
from transformers import (
AutoProcessor,
)
from text_generation_server.models.custom_modeling.llava_next import (
LlavaNextForConditionalGeneration,
)
from text_generation_server.models.vlm_causal_lm import VlmCausalLM
class LlavaNext(VlmCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.processor = AutoProcessor.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
super().__init__(
model_cls=LlavaNextForConditionalGeneration,
model_id=model_id,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
def get_layer_config(self, model) -> Tuple[int, int, int]:
return (
len(model.language_model.model.layers),
model.language_model.model.num_key_value_heads,
model.language_model.model.head_size,
)
def max_past(self) -> Optional[int]:
return getattr(self.model.language_model, "max_past", None)
server/text_generation_server/models/mpt.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from pathlib import Path
from typing import Optional, Type
from opentelemetry import trace
from transformers import AutoTokenizer, PretrainedConfig, PreTrainedTokenizerBase
from huggingface_hub import hf_hub_download
import json
from text_generation_server.models import CausalLM
from text_generation_server.models.causal_lm import CausalLMBatch
from text_generation_server.pb import generate_pb2
from text_generation_server.models.custom_modeling.mpt_modeling import (
MPTForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
tracer = trace.get_tracer(__name__)
class MPTCausalLMBatch(CausalLMBatch):
@classmethod
def from_pb(
cls,
pb: generate_pb2.Batch,
tokenizer: PreTrainedTokenizerBase,
dtype: torch.dtype,
device: torch.device,
) -> "CausalLMBatch":
batch = super().from_pb(pb=pb, tokenizer=tokenizer, dtype=dtype, device=device)
batch.keys_head_dim_last = False
return batch
class MPTSharded(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
tokenizer.pad_token = tokenizer.eos_token
# If model_id is a local path, load the file directly
local_path = Path(model_id, "config.json")
if local_path.exists():
filename = str(local_path.resolve())
else:
filename = hf_hub_download(
model_id, revision=revision, filename="config.json"
)
with open(filename, "r") as f:
config = json.load(f)
config = PretrainedConfig(**config)
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
config.quantize = quantize
model = MPTForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=False,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
@property
def batch_type(self) -> Type[CausalLMBatch]:
return MPTCausalLMBatch
server/text_generation_server/models/opt.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from typing import Optional
from transformers import (
AutoTokenizer,
AutoConfig,
)
from text_generation_server.models.custom_modeling.opt_modeling import OPTForCausalLM
from text_generation_server.models import CausalLM
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
class OPTSharded(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = AutoConfig.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
)
config.quantize = quantize
config.speculator = speculator
tokenizer.pad_token_id = config.pad_token_id
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames, device=device, dtype=dtype, process_group=self.process_group
)
if config.quantize in ["gptq", "marlin"]:
weights._set_gptq_params(model_id, revision)
model = OPTForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
):
outputs, speculative_logits = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
use_cache=True,
)
return outputs.logits, speculative_logits, outputs.past_key_values
server/text_generation_server/models/pali_gemma.py
View file @ fb2f74e2
...@@ -74,45 +74,3 @@ class PaliGemmaBatch(VlmCausalLMBatch): ...@@ -74,45 +74,3 @@ class PaliGemmaBatch(VlmCausalLMBatch):
else: else:
image_inputs = None image_inputs = None
return batch_tokenized_inputs, image_inputs return batch_tokenized_inputs, image_inputs
class PaliGemma(VlmCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.processor = AutoProcessor.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
)
super().__init__(
config_cls=AutoConfig,
model_cls=PaliGemmaForConditionalGeneration,
model_id=model_id,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
@property
def batch_type(self):
return PaliGemmaBatch
def get_layer_config(self, model) -> Tuple[int, int, int]:
return (
len(model.text_model.model.layers),
model.text_model.model.num_key_value_heads,
model.text_model.model.head_size,
)
def max_past(self) -> Optional[int]:
return getattr(self.model.text_model, "max_past", None)
server/text_generation_server/models/phi.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from transformers import AutoConfig, AutoTokenizer
from typing import Optional, List, Tuple
from text_generation_server.models import CausalLM
from text_generation_server.models.custom_modeling.phi_modeling import (
PhiConfig,
PhiForCausalLM,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
class Phi(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, _rank, _world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device("cuda")
dtype = torch.float16 if dtype is None else dtype
else:
if quantize:
raise ValueError("quantization is not available on CPU")
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
config = PhiConfig.from_pretrained(
model_id, revision=revision, trust_remote_code=trust_remote_code
)
tokenizer.bos_token_id = config.bos_token_id
tokenizer.eos_token_id = config.eos_token_id
tokenizer.pad_token = tokenizer.eos_token
config.quantize = quantize
config.speculator = speculator
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(filenames, device, dtype, process_group=self.process_group)
model = PhiForCausalLM(config, weights)
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
)
server/text_generation_server/models/rw.py deleted 100644 → 0
View file @ c6bcadf8
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from typing import List, Optional, Tuple
from text_generation_server.models import CausalLM
class RW(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
if speculator:
raise RuntimeError("Medusa decoding is not enabled for AutoModel")
if torch.cuda.is_available():
device = torch.device("cuda")
dtype = torch.float16 if dtype is None else dtype
else:
if quantize:
raise ValueError("quantization is not available on CPU")
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
model = AutoModelForCausalLM.from_pretrained(
model_id,
revision=revision,
torch_dtype=dtype,
device_map=(
"auto"
if torch.cuda.is_available() and torch.cuda.device_count() > 1
else None
),
load_in_8bit=quantize == "bitsandbytes",
trust_remote_code=trust_remote_code,
)
if torch.cuda.is_available() and torch.cuda.device_count() == 1:
model = model.cuda()
if tokenizer.pad_token_id is None:
if model.config.pad_token_id is not None:
tokenizer.pad_token_id = model.config.pad_token_id
elif model.config.eos_token_id is not None:
tokenizer.pad_token_id = model.config.eos_token_id
elif tokenizer.eos_token_id is not None:
tokenizer.pad_token_id = tokenizer.eos_token_id
else:
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
)
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
):
# Model Forward
outputs, speculative_logits = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
use_cache=True,
)
return outputs.logits, speculative_logits, outputs.past_key_values
server/text_generation_server/models/santacoder.py deleted 100644 → 0
View file @ c6bcadf8
import torch
import torch.distributed
from typing import Optional, List
from transformers import AutoTokenizer, AutoModelForCausalLM
from text_generation_server.models import CausalLM
FIM_PREFIX = "<fim-prefix>"
FIM_MIDDLE = "<fim-middle>"
FIM_SUFFIX = "<fim-suffix>"
FIM_PAD = "<fim-pad>"
EOD = "<|endoftext|>"
class SantaCoder(CausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
if torch.cuda.is_available():
device = torch.device("cuda")
dtype = torch.float16 if dtype is None else dtype
else:
if quantize:
raise ValueError("quantization is not available on CPU")
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
tokenizer.add_special_tokens(
{
"additional_special_tokens": [
EOD,
FIM_PREFIX,
FIM_MIDDLE,
FIM_SUFFIX,
FIM_PAD,
],
"pad_token": EOD,
}
)
with device:
model = AutoModelForCausalLM.from_pretrained(
model_id,
revision=revision,
torch_dtype=dtype,
load_in_8bit=quantize == "bitsandbytes",
trust_remote_code=trust_remote_code,
)
super(CausalLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
)
def decode(self, generated_ids: List[int]) -> str:
# Do not skip special tokens as they are used for custom parsing rules of the generated text
return self.tokenizer.decode(
generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment