huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-22 15:32:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add support for other vlm

Browse Source
This commit is contained in:
Mohit Sharma 2025-03-21 11:22:12 +00:00
parent d8d09f9c7b
commit ac6fc70c75
2 changed files with 361 additions and 161 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

124
server/text_generation_server/models/__init__.py
View File

@ -208,6 +208,8 @@ try:
)
from text_generation_server.models.transformers_flash_vlm import (
TransformersFlashVlmCausalLM,
TransformersQwen2VlmCausalLM,
TransformersGemma3VlmCausalLM,
)
except ImportError as e:
log_master(logger.warning, f"Could not import Flash Transformers Backend: {e}")
@ -1161,7 +1163,6 @@ def get_model(
)
elif model_type == GEMMA3:
if FLASH_ATTENTION:
# TODO: Use VlmCausalLM when image support is added.
return VlmCausalLM(
model_id=model_id,
model_class=Gemma3ForConditionalGeneration,
@ -1179,9 +1180,11 @@ def get_model(
lora_adapter_ids=lora_adapter_ids,
)
elif FLASH_TRANSFORMERS_BACKEND:
return TransformersFlashVlmCausalLM.fallback(
from transformers import Gemma3ForConditionalGeneration as Gemma3Model
return TransformersGemma3VlmCausalLM.fallback(
model_id,
# AutoModelForConditionalGeneration,
Gemma3Model,
revision,
quantize=quantize,
speculator=speculator,
@ -1490,15 +1493,7 @@ def get_model(
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
if model_type == QWEN2_VL:
return TransformersFlashVlmCausalLM.fallback(
model_id,
# AutoModelForConditionalGeneration,
revision,
quantize=quantize,
speculator=speculator,
dtype=torch.bfloat16,
trust_remote_code=trust_remote_code,
)
if FLASH_ATTENTION:
return VlmCausalLM(
model_id=model_id,
model_class=Qwen2VLForConditionalGeneration,
@ -1511,7 +1506,20 @@ def get_model(
trust_remote_code=trust_remote_code,
lora_adapter_ids=lora_adapter_ids,
)
elif FLASH_TRANSFORMERS_BACKEND:
from transformers import Qwen2VLForConditionalGeneration as Qwen2VLModel
return TransformersQwen2VlmCausalLM.fallback(
model_id,
Qwen2VLModel,
revision,
quantize=quantize,
speculator=speculator,
dtype=torch.bfloat16,
trust_remote_code=trust_remote_code,
)
if model_type == QWEN2_5_VL:
if FLASH_ATTENTION:
return VlmCausalLM(
model_id=model_id,
model_class=Qwen2_5VLForConditionalGeneration,
@ -1526,6 +1534,19 @@ def get_model(
config_class=Qwen2_5_VLConfig,
processor_class=Qwen2_5_VLProcessor,
)
elif FLASH_TRANSFORMERS_BACKEND:
return TransformersQwen2VlmCausalLM.fallback(
model_id,
Qwen2VLModel,
revision,
quantize=quantize,
speculator=speculator,
dtype=torch.bfloat16,
trust_remote_code=trust_remote_code,
config_class=Qwen2_5_VLConfig,
processor_class=Qwen2_5_VLProcessor,
)
if model_type == MLLAMA:
if FLASH_ATTENTION:
return MllamaCausalLM(
@ -1540,6 +1561,19 @@ def get_model(
trust_remote_code=trust_remote_code,
lora_adapter_ids=lora_adapter_ids,
)
elif FLASH_TRANSFORMERS_BACKEND:
from transformers import MllamaForConditionalGeneration as MllamaModel
return TransformersFlashVlmCausalLM.fallback(
model_id,
MllamaModel,
revision,
quantize=quantize,
speculator=speculator,
dtype=torch.bfloat16,
trust_remote_code=trust_remote_code,
batch_class=MllamaCausalLMBatch,
)
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Mllama"))
if model_type == IDEFICS2:
@ -1558,6 +1592,19 @@ def get_model(
# VRAM usage.
processor_kwargs={"size": {"longest_edge": 448, "shortest_edge": 378}},
)
elif FLASH_TRANSFORMERS_BACKEND:
from transformers import Idefics2ForConditionalGeneration as Idefics2Model
return TransformersFlashVlmCausalLM.fallback(
model_id,
Idefics2Model,
revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
processor_kwargs={"size": {"longest_edge": 448, "shortest_edge": 378}},
)
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
if model_type == IDEFICS3:
@ -1576,20 +1623,23 @@ def get_model(
# VRAM usage.
processor_kwargs={"size": {"longest_edge": 1456}},
)
elif FLASH_TRANSFORMERS_BACKEND:
from transformers import Idefics3ForConditionalGeneration as Idefics3Model
return TransformersFlashVlmCausalLM.fallback(
model_id,
Idefics3Model,
revision,
quantize=quantize,
speculator=speculator,
dtype=torch.bfloat16,
trust_remote_code=trust_remote_code,
processor_kwargs={"size": {"longest_edge": 1456}},
)
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
if model_type == PALIGEMMA:
# return TransformersFlashVlmCausalLM.fallback(
# model_id,
# # AutoModelForConditionalGeneration,
# revision,
# quantize=quantize,
# speculator=speculator,
# dtype=torch.bfloat16,
# trust_remote_code=trust_remote_code,
# batch_class=PaliGemmaBatch,
# )
# if FLASH_ATTENTION:
if FLASH_ATTENTION:
return VlmCausalLM(
model_id=model_id,
model_class=PaliGemmaForConditionalGeneration,
@ -1604,9 +1654,21 @@ def get_model(
lora_adapter_ids=lora_adapter_ids,
batch_class=PaliGemmaBatch,
)
# else:
# raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
elif FLASH_TRANSFORMERS_BACKEND:
from transformers import PaliGemmaForConditionalGeneration as PaliGemmaModel
return TransformersFlashVlmCausalLM.fallback(
model_id,
PaliGemmaModel,
revision,
quantize=quantize,
speculator=speculator,
dtype=torch.bfloat16,
trust_remote_code=trust_remote_code,
batch_class=PaliGemmaBatch,
)
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("PaliGemma"))
if model_type == LLAVA_NEXT:
if FLASH_ATTENTION:
return VlmCausalLM(
@ -1619,6 +1681,18 @@ def get_model(
kv_cache_dtype=kv_cache_dtype,
trust_remote_code=trust_remote_code,
)
elif FLASH_TRANSFORMERS_BACKEND:
from transformers import LlavaNextForConditionalGeneration as LlavaNextModel
return TransformersFlashVlmCausalLM.fallback(
model_id,
LlavaNextModel,
revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("LlavaNext"))

320
server/text_generation_server/models/transformers_flash_vlm.py
View File

@ -17,6 +17,15 @@ import torch.nn.functional as F
tracer = trace.get_tracer(__name__)
# The base TP plan of these models has replicated q/k/v. This means that each process will see the full states,
# hence we should not divide the number of heads by the world size. This is a known waste of VRAM (the cache
# will be fully replicated on each process) and GPU communication (additional all-gather operations), however due
# to internal constraints it was not (yet?) possible to circumvent
REPLICATED_ATTENTION_MODELS = [
"olmo2",
"phi3",
]
def tgi_flash_attention_forward(
module,
@ -35,17 +44,13 @@ def tgi_flash_attention_forward(
softmax_scale: Optional[float] = None,
sliding_window: Optional[int] = None,
softcap: Optional[float] = None,
use_sdpa: Optional[bool] = False,
**kwargs, # This is needed to "absorb" other args passed by Transformers modeling
):
from loguru import logger
logger.info("Using TGI Flash Attention")
# from pdb import set_trace; set_trace()
kv_cache = kv_cache[module.layer_idx]
query_states = query_states.transpose(1, 2).squeeze(dim=0)
key_states = key_states.transpose(1, 2).squeeze(dim=0)
value_states = value_states.transpose(1, 2).squeeze(dim=0)
# from pdb import set_trace; set_trace()
# Take care of updating the cache in-place
kv_cache.store(key=key_states, value=value_states, slots=slots, kv_scales=kv_scales)
@ -55,10 +60,8 @@ def tgi_flash_attention_forward(
sliding_window = -1 if sliding_window is None else sliding_window
# if module.layer_idx == 0:
# from pdb import set_trace; set_trace()
if cu_seqlen_prefill is not None:
attention_mask = None
if attention_mask is None:
if not use_sdpa:
attn_output = attention(
query=query_states,
key=key_states,
@ -72,9 +75,6 @@ def tgi_flash_attention_forward(
softcap=softcap,
)
else:
from loguru import logger
logger.info("uSING FLASH ATTENTION")
lengths = cu_seqlen_prefill[1:] - cu_seqlen_prefill[:-1]
max_length = max(lengths)
attention_mask = attention_mask[:, :, :, :max_length]
@ -135,34 +135,62 @@ def tgi_flash_attention_forward(
transformers.modeling_utils.ALL_ATTENTION_FUNCTIONS["tgi"] = tgi_flash_attention_forward
# Siglip
transformers.models.siglip.modeling_siglip.SIGLIP_ATTENTION_CLASSES["tgi"] = (
transformers.models.siglip.modeling_siglip.SIGLIP_ATTENTION_CLASSES["sdpa"]
)
# transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_ATTENTION_CLASSES["tgi"] = transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_VISION_ATTENTION_CLASSES["eager"]
# transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_ATTENTION_CLASSES["tgi"] = transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_VISION_ATTENTION_CLASSES["eager"]
transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_ATTENTION_CLASSES["tgi"] = (
tgi_flash_attention_forward
)
# Qwen2VL
transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_VISION_ATTENTION_CLASSES[
"tgi"
] = transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_VISION_ATTENTION_CLASSES[
"eager"
]
# This needs to be patched in transformers to use ALL_ATTENTION_FUNCTIONS
# transformers.models.qwen2_vl.modeling_qwen2_vl.QWEN2_VL_ATTENTION_CLASSES["tgi"] = tgi_flash_attention_forward
# The base TP plan of these models has replicated q/k/v. This means that each process will see the full states,
# hence we should not divide the number of heads by the world size. This is a known waste of VRAM (the cache
# will be fully replicated on each process) and GPU communication (additional all-gather operations), however due
# to internal constraints it was not (yet?) possible to circumvent
REPLICATED_ATTENTION_MODELS = [
"olmo2",
"phi3",
# Idefics2
transformers.models.idefics2.modeling_idefics2.IDEFICS_VISION_ATTENTION_CLASSES[
"tgi"
] = transformers.models.idefics2.modeling_idefics2.IDEFICS_VISION_ATTENTION_CLASSES[
"eager"
]
transformers.models.idefics2.modeling_idefics2.IDEFICS2_PERCEIVER_ATTENTION_CLASSES[
"tgi"
] = transformers.models.idefics2.modeling_idefics2.IDEFICS2_PERCEIVER_ATTENTION_CLASSES[
"eager"
]
# Idefics3
transformers.models.idefics3.modeling_idefics3.IDEFICS_VISION_ATTENTION_CLASSES[
"tgi"
] = transformers.models.idefics3.modeling_idefics3.IDEFICS_VISION_ATTENTION_CLASSES[
"eager"
]
# Clip
transformers.models.clip.modeling_clip.CLIP_ATTENTION_CLASSES["tgi"] = (
transformers.models.clip.modeling_clip.CLIP_ATTENTION_CLASSES["sdpa"]
)
# Mllama
transformers.models.mllama.modeling_mllama.MLLAMA_VISION_ATTENTION_CLASSES["tgi"] = (
transformers.models.mllama.modeling_mllama.MLLAMA_VISION_ATTENTION_CLASSES["eager"]
)
# This needs to be patched in transformers to use ALL_ATTENTION_FUNCTIONS
# transformers.models.mllama.modeling_mllama.MLLAMA_TEXT_ATTENTION_CLASSES["tgi"] = tgi_flash_attention_forward
# transformers.models.mllama.modeling_mllama.MLLAMA_CROSS_ATTENTION_CLASSES["tgi"] = tgi_cross_attention_forward
# TODO: implement
# tgi_cross_attention_forward
class TransformersFlashVlmCausalLM(VlmCausalLM):
def __init__(
self,
model_id: str,
model_class,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
@ -175,10 +203,10 @@ class TransformersFlashVlmCausalLM(VlmCausalLM):
kv_cache_dtype: Optional[torch.dtype] = None,
batch_class=VlmCausalLMBatch,
):
# # from pdb import set_trace; set_trace()
self.batch_class = VlmCausalLMBatch
self.batch_class = batch_class
self.quantize = quantize
self.process_group, rank, world_size = initialize_torch_distributed()
self.dtype = dtype
if speculator:
raise RuntimeError("Speculator decoding is not enabled for AutoModel")
@ -204,16 +232,15 @@ class TransformersFlashVlmCausalLM(VlmCausalLM):
if processor_kwargs is None:
processor_kwargs = {}
# processor_kwargs={"size": {"longest_edge": 448, "shortest_edge": 378}}
self.processor = processor_class.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
**processor_kwargs,
)
from transformers import Qwen2VLForConditionalGeneration
model = Qwen2VLForConditionalGeneration.from_pretrained(
model = model_class.from_pretrained(
model_id,
revision=revision,
torch_dtype=dtype,
@ -318,6 +345,100 @@ class TransformersFlashVlmCausalLM(VlmCausalLM):
torch.distributed.barrier(group=self.process_group)
def get_position_ids(self, input_ids, image_grid_thw, position_ids):
return position_ids
def pre_process_inputs(self, input_ids, position_ids, cu_seqlen_prefill):
return {
"input_ids": input_ids.unsqueeze(0),
"position_ids": position_ids.unsqueeze(0),
}
def post_process_outputs(self, logits, lm_head_indices):
return logits.squeeze(dim=0)
@classmethod
def fallback(
cls,
model_id: str,
model_class,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
batch_class: Optional[type] = VlmCausalLMBatch,
processor_kwargs: Optional[dict] = None,
):
return cls(
model_id=model_id,
model_class=model_class,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
batch_class=batch_class,
processor_kwargs=processor_kwargs,
)
def _model_forward(
self,
input_ids: torch.Tensor,
position_ids: torch.Tensor,
cu_seqlen_prefill: Optional[torch.Tensor],
kv_cache: List[KVCache],
block_tables: torch.Tensor,
slots: torch.Tensor,
seqlen: Seqlen,
max_s: int,
lm_head_indices: Optional[torch.Tensor],
prefill_cache_indices=None, # not used, but passed to match original signature
adapter_data=None, # not supported, but passed to match original signature
pixel_values: torch.FloatTensor = None,
image_grid_thw: Optional[torch.LongTensor] = None,
pixel_attention_mask=None,
image_sizes: Optional[torch.LongTensor] = None,
):
# A value of `None` (i.e. no logit slicing) translates to `0` in Transformers
logits_to_keep = lm_head_indices if lm_head_indices is not None else 0
inputs = self.pre_process_inputs(
input_ids=input_ids,
position_ids=position_ids,
cu_seqlen_prefill=cu_seqlen_prefill,
)
# This is equivalent to `self.model.forward`, see the monkey patch in __init__
logits = self.model.original_forward(
input_ids=inputs["input_ids"],
position_ids=inputs["position_ids"],
past_key_values=None, # we use self.kv_cache instead of transformers cache object
use_cache=False, # we use self.kv_cache instead of transformers cache object
logits_to_keep=logits_to_keep,
return_dict=True,
cu_seqlen_prefill=cu_seqlen_prefill,
kv_cache=kv_cache,
block_tables=block_tables,
slots=slots,
seqlen=seqlen,
max_s=max_s,
kv_head_mapping=self.kv_head_mapping,
kv_scales=self.kv_scales,
pixel_values=pixel_values,
pixel_attention_mask=pixel_attention_mask,
image_sizes=image_sizes,
image_grid_thw=image_grid_thw,
attention_mask=inputs.get("attention_mask", None),
use_sdpa=inputs.get("use_sdpa", False),
).logits
logits = self.post_process_outputs(logits, lm_head_indices)
return logits, None
class TransformersQwen2VlmCausalLM(TransformersFlashVlmCausalLM):
def get_position_ids(self, input_ids: torch.Tensor, image_grid_thw: torch.Tensor):
if image_grid_thw is None:
return (
@ -391,77 +512,82 @@ class TransformersFlashVlmCausalLM(VlmCausalLM):
)
return position_ids
@classmethod
def fallback(
cls,
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,
batch_class: Optional[type] = VlmCausalLMBatch,
):
return cls(
model_id=model_id,
revision=revision,
quantize=quantize,
speculator=speculator,
dtype=dtype,
trust_remote_code=trust_remote_code,
batch_class=batch_class,
def post_process_outputs(self, logits, lm_head_indices):
return logits.squeeze(dim=0)[lm_head_indices].unsqueeze(0)
def pre_process_inputs(self, input_ids, position_ids, cu_seqlen_prefill):
input_ids = input_ids.unsqueeze(0)
position_ids = position_ids.transpose(0, 1).unsqueeze(1)
return {"input_ids": input_ids, "position_ids": position_ids}
class TransformersGemma3VlmCausalLM(TransformersFlashVlmCausalLM):
def get_attention_mask(self, input_ids, cu_seqlen_prefill):
device = input_ids.device
dtype = self.dtype
min_dtype = torch.finfo(dtype).min
lengths = (cu_seqlen_prefill[1:] - cu_seqlen_prefill[:-1]).tolist()
batch_size = len(lengths)
sequence_length = max(lengths)
target_length = sequence_length
# Create the padding mask from the computed lengths.
# pad_mask: [batch, sequence_length] where True indicates valid tokens.
seq_range = torch.arange(sequence_length, device=device).unsqueeze(0)
lengths_tensor = torch.tensor(lengths, device=device).unsqueeze(1)
pad_mask = seq_range < lengths_tensor # shape: [batch, sequence_length]
# Build the base causal mask (for non-image tokens):
causal_mask = torch.tril(
torch.ones(
(sequence_length, sequence_length), dtype=torch.bool, device=device
)
def _model_forward(
self,
input_ids: torch.Tensor,
position_ids: torch.Tensor,
cu_seqlen_prefill: Optional[torch.Tensor],
kv_cache: List[KVCache],
block_tables: torch.Tensor,
slots: torch.Tensor,
seqlen: Seqlen,
max_s: int,
lm_head_indices: Optional[torch.Tensor],
prefill_cache_indices=None, # not used, but passed to match original signature
adapter_data=None, # not supported, but passed to match original signature
pixel_values: torch.FloatTensor = None,
image_grid_thw: Optional[torch.LongTensor] = None,
pixel_attention_mask=None,
image_sizes: Optional[torch.LongTensor] = None,
):
# from pdb import set_trace; set_trace()
# A value of `None` (i.e. no logit slicing) translates to `0` in Transformers
logits_to_keep = lm_head_indices if lm_head_indices is not None else 0
# This is equivalent to `self.model.forward`, see the monkey patch in __init__
logits = (
self.model.original_forward(
input_ids=input_ids.unsqueeze(0), # expand dim to fit Transformers
position_ids=position_ids.transpose(0, 1).unsqueeze(
)
base_mask = pad_mask.unsqueeze(2) & pad_mask.unsqueeze(
1
), # expand dim to fit Transformers
past_key_values=None, # we use self.kv_cache instead of transformers cache object
use_cache=False, # we use self.kv_cache instead of transformers cache object
logits_to_keep=logits_to_keep,
return_dict=True,
cu_seqlen_prefill=cu_seqlen_prefill,
kv_cache=kv_cache,
block_tables=block_tables,
slots=slots,
seqlen=seqlen,
max_s=max_s,
kv_head_mapping=self.kv_head_mapping,
kv_scales=self.kv_scales,
pixel_values=pixel_values,
pixel_attention_mask=pixel_attention_mask,
image_sizes=image_sizes,
image_grid_thw=image_grid_thw,
)
.logits.squeeze(dim=0)[lm_head_indices]
.unsqueeze(0)
) # [batch, sequence_length, sequence_length]
base_mask = base_mask & causal_mask.unsqueeze(0) # apply causal constraint
image_token_mask = (input_ids == self.config.image_token_index).to(
input_ids.device
)
# from pdb import set_trace; set_trace()
image_token_mask = torch.nn.utils.rnn.pad_sequence(
torch.split(image_token_mask, lengths), batch_first=True, padding_value=0
)
bidirectional_mask = image_token_mask.unsqueeze(2) & image_token_mask.unsqueeze(
1
)
return logits, None
# Combine the causal base mask and the bidirectional mask.
combined_mask = torch.logical_or(
base_mask.unsqueeze(1), bidirectional_mask.unsqueeze(1)
).to(device)
# combined_mask now has shape [batch, 1, sequence_length, sequence_length]
full_attention_mask = torch.zeros(
(batch_size, 1, sequence_length, target_length),
device=device,
dtype=torch.bool,
)
full_attention_mask[:, :, :, :sequence_length] = combined_mask
final_attention_mask = torch.where(full_attention_mask, 0, min_dtype).to(device)
return final_attention_mask
def pre_process_inputs(self, input_ids, position_ids, cu_seqlen_prefill):
inputs = {
"input_ids": input_ids.unsqueeze(0),
"position_ids": position_ids.unsqueeze(0),
}
if cu_seqlen_prefill is not None:
attention_mask = self.get_attention_mask(
input_ids.squeeze(0), cu_seqlen_prefill
)
inputs["attention_mask"] = attention_mask
inputs["use_sdpa"] = True
return inputs