huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-11-18 23:15:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

Multi modality fix (#3283)

Browse Source 
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
This commit is contained in:
Wang, Yi 2025-08-19 15:36:36 +08:00 committed by GitHub
parent 6a2fa83540
commit 5284b5c654
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 107 additions and 61 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
server/text_generation_server/models/custom_modeling/mllama.py
View File

@ -710,22 +710,29 @@ class MllamaTextCrossAttention(nn.Module):
# ) # )
if SYSTEM == "ipex": if SYSTEM == "ipex":
attn_output = torch.empty_like(query_states) attn_output = torch.empty_like(query_states)
if query_states.device.type == "xpu":
ipex.llm.functional.varlen_attention( ipex.llm.functional.varlen_attention(
( query_states.contiguous(),
query_states.contiguous() key_states.contiguous(),
if query_states.device.type == "xpu" value_states.contiguous(),
else query_states attn_output,
), cu_seqlen_q,
( cu_seqlen_k,
key_states.contiguous() None,
if key_states.device.type == "xpu" max_q,
else key_states max_k,
), 0.0,
( self.softmax_scale,
value_states.contiguous() False,
if value_states.device.type == "xpu" causal,
else value_states False,
), None,
)
else:
ipex.llm.functional.varlen_attention(
query_states,
key_states,
value_states,
attn_output, attn_output,
cu_seqlen_q, cu_seqlen_q,
cu_seqlen_k, cu_seqlen_k,

25
server/text_generation_server/models/custom_modeling/qwen2_5_vl.py
View File

@ -460,10 +460,29 @@ class Qwen2_5VLAttention(nn.Module):
# execute flash attention # execute flash attention
if SYSTEM == "ipex": if SYSTEM == "ipex":
attn_output = torch.empty_like(query) attn_output = torch.empty_like(query)
if query.device.type == "xpu":
ipex.llm.functional.varlen_attention( ipex.llm.functional.varlen_attention(
(query.contiguous() if query.device.type == "xpu" else query), query.contiguous(),
(key.contiguous() if key.device.type == "xpu" else key), key.contiguous(),
(value.contiguous() if value.device.type == "xpu" else value), value.contiguous(),
attn_output,
cu_seqlens,
cu_seqlens,
None,
max_seqlen,
max_seqlen,
0.0,
self.softmax_scale,
False,
causal,
False,
None,
)
else:
ipex.llm.functional.varlen_attention(
query,
key,
value,
attn_output, attn_output,
cu_seqlens, cu_seqlens,
cu_seqlens, cu_seqlens,

25
server/text_generation_server/models/custom_modeling/qwen2_vl.py
View File

@ -130,10 +130,29 @@ class Qwen2VLAttention(nn.Module):
# execute flash attention # execute flash attention
if SYSTEM == "ipex": if SYSTEM == "ipex":
attn_output = torch.empty_like(query) attn_output = torch.empty_like(query)
if query.device.type == "xpu":
ipex.llm.functional.varlen_attention( ipex.llm.functional.varlen_attention(
(query.contiguous() if query.device.type == "xpu" else query), query.contiguous(),
(key.contiguous() if key.device.type == "xpu" else key), key.contiguous(),
(value.contiguous() if value.device.type == "xpu" else value), value.contiguous(),
attn_output,
cu_seqlens,
cu_seqlens,
None,
max_seqlen,
max_seqlen,
0.0,
self.softmax_scale,
False,
causal,
False,
None,
)
else:
ipex.llm.functional.varlen_attention(
query,
key,
value,
attn_output, attn_output,
cu_seqlens, cu_seqlens,
cu_seqlens, cu_seqlens,

3
server/text_generation_server/models/mllama_causal_lm.py
View File

@ -59,7 +59,7 @@ class MllamaCausalLMBatch(VlmCausalLMBatch):
@tracer.start_as_current_span("filter") @tracer.start_as_current_span("filter")
def filter(self, request_ids: List[int]): def filter(self, request_ids: List[int]):
assert self.image_indices is not None assert self.image_indices is not None
batch = super().filter(request_ids) batch = super(VlmCausalLMBatch, self).filter(request_ids)
assert self.image_indices is not None assert self.image_indices is not None
indices = [] indices = []
for i, request_id in enumerate(request_ids): for i, request_id in enumerate(request_ids):
@ -85,6 +85,7 @@ class MllamaCausalLMBatch(VlmCausalLMBatch):
] ]
else: else:
batch.cross_attention_states = None batch.cross_attention_states = None
batch.pixel_values = None
return batch return batch
@classmethod @classmethod