huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-09 03:14:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

xpu add alibi_scope input in varlen_attention in ipex 2.7 while cpu does not. so split the case.

Browse Source 
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
This commit is contained in:
Wang, Yi A 2025-06-29 22:15:12 -07:00
parent 9f38d93051
commit d05a5c3f0a
3 changed files with 105 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
server/text_generation_server/models/custom_modeling/mllama.py
View File

@ -710,34 +710,41 @@ class MllamaTextCrossAttention(nn.Module):
# ) # )
if SYSTEM == "ipex": if SYSTEM == "ipex":
attn_output = torch.empty_like(query_states) attn_output = torch.empty_like(query_states)
ipex.llm.functional.varlen_attention( if query_states.device.type == "xpu":
( ipex.llm.functional.varlen_attention(
query_states.contiguous() query_states.contiguous(),
if query_states.device.type == "xpu" key_states.contiguous(),
else query_states value_states.contiguous(),
), attn_output,
( cu_seqlen_q,
key_states.contiguous() cu_seqlen_k,
if key_states.device.type == "xpu" None,
else key_states max_q,
), max_k,
( 0.0,
value_states.contiguous() self.softmax_scale,
if value_states.device.type == "xpu" False,
else value_states causal,
), False,
attn_output, None,
cu_seqlen_q, )
cu_seqlen_k, else:
max_q, ipex.llm.functional.varlen_attention(
max_k, query_states,
0.0, key_states,
self.softmax_scale, value_states,
False, attn_output,
causal, cu_seqlen_q,
False, cu_seqlen_k,
None, max_q,
) max_k,
0.0,
self.softmax_scale,
False,
causal,
False,
None,
)
else: else:
attn_output = flash_attn_2_cuda.varlen_fwd( attn_output = flash_attn_2_cuda.varlen_fwd(
query_states, query_states,

51
server/text_generation_server/models/custom_modeling/qwen2_5_vl.py
View File

@ -460,22 +460,41 @@ 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)
ipex.llm.functional.varlen_attention( if query.device.dtype == "xpu":
(query.contiguous() if query.device.type == "xpu" else query), ipex.llm.functional.varlen_attention(
(key.contiguous() if key.device.type == "xpu" else key), query.contiguous(),
(value.contiguous() if value.device.type == "xpu" else value), key.contiguous(),
attn_output, value.contiguous(),
cu_seqlens, attn_output,
cu_seqlens, cu_seqlens,
max_seqlen, cu_seqlens,
max_seqlen, None,
0.0, max_seqlen,
self.softmax_scale, max_seqlen,
False, 0.0,
causal, self.softmax_scale,
False, False,
None, causal,
) False,
None,
)
else:
ipex.llm.functional.varlen_attention(
query,
key,
value,
attn_output,
cu_seqlens,
cu_seqlens,
max_seqlen,
max_seqlen,
0.0,
self.softmax_scale,
False,
causal,
False,
None,
)
else: else:
attn_output = flash_attn_2_cuda.varlen_fwd( attn_output = flash_attn_2_cuda.varlen_fwd(
query, query,

51
server/text_generation_server/models/custom_modeling/qwen2_vl.py
View File

@ -130,22 +130,41 @@ 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)
ipex.llm.functional.varlen_attention( if query.device.type == "xpu":
(query.contiguous() if query.device.type == "xpu" else query), ipex.llm.functional.varlen_attention(
(key.contiguous() if key.device.type == "xpu" else key), query.contiguous(),
(value.contiguous() if value.device.type == "xpu" else value), key.contiguous(),
attn_output, value.contiguous(),
cu_seqlens, attn_output,
cu_seqlens, cu_seqlens,
max_seqlen, cu_seqlens,
max_seqlen, None,
0.0, max_seqlen,
self.softmax_scale, max_seqlen,
False, 0.0,
causal, self.softmax_scale,
False, False,
None, causal,
) False,
None,
)
else:
ipex.llm.functional.varlen_attention(
query,
key,
value,
attn_output,
cu_seqlens,
cu_seqlens,
max_seqlen,
max_seqlen,
0.0,
self.softmax_scale,
False,
causal,
False,
None,
)
else: else:
attn_output = flash_attn_2_cuda.varlen_fwd( attn_output = flash_attn_2_cuda.varlen_fwd(
query, query,