xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 02:05:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Re-enable FlashInfer for Llama4 on Blackwell in e2e fusion tests (#28966)

Browse Source 
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: ProExpertProg <11367180+ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
Copilot 2025-11-19 13:32:19 -05:00 committed by GitHub
parent 0c80efd94f
commit 61728cd1df
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 6 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.buildkite/test-pipeline.yaml
View File

@ -930,6 +930,8 @@ steps:
- csrc/quantization/fp4/ - csrc/quantization/fp4/
- vllm/model_executor/layers/quantization/utils/flashinfer_utils.py - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
- vllm/v1/attention/backends/flashinfer.py - vllm/v1/attention/backends/flashinfer.py
- vllm/v1/worker/
- vllm/v1/cudagraph_dispatcher.py
- vllm/compilation/ - vllm/compilation/
# can affect pattern matching # can affect pattern matching
- vllm/model_executor/layers/layernorm.py - vllm/model_executor/layers/layernorm.py

12
tests/compile/distributed/test_fusions_e2e.py
View File

@ -47,12 +47,8 @@ if current_platform.is_cuda():
ModelBackendTestCase( ModelBackendTestCase(
# Use smaller model for L40s in CI # Use smaller model for L40s in CI
model_name="RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8", model_name="RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8",
# TODO while llama4 is broken, use FLASHINFER for llama3 on Blackwell
# so FI attention+fp8_quant is at least tested once
model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"), model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"),
backend=AttentionBackendEnum.FLASHINFER backend=AttentionBackendEnum.TRITON_ATTN,
if is_blackwell()
else AttentionBackendEnum.TRITON_ATTN,
matches=Matches( matches=Matches(
attention_fusion=32, attention_fusion=32,
allreduce_fusion=65, allreduce_fusion=65,
@ -65,9 +61,9 @@ if current_platform.is_cuda():
model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"), model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"),
# TODO FlashInfer attn broken on Hopper with kvcache=fp8: # TODO FlashInfer attn broken on Hopper with kvcache=fp8:
# https://github.com/vllm-project/vllm/issues/28568 # https://github.com/vllm-project/vllm/issues/28568
# TODO FlashInfer attn broken on Blackwell for llama4: backend=AttentionBackendEnum.FLASHINFER
# https://github.com/vllm-project/vllm/issues/28604 if is_blackwell()
backend=AttentionBackendEnum.TRITON_ATTN, else AttentionBackendEnum.TRITON_ATTN,
matches=Matches( matches=Matches(
attention_fusion=48, attention_fusion=48,
allreduce_fusion=96, allreduce_fusion=96,