[UX] Integrate DeepGEMM into vLLM wheel via CMake (#37980)

Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Michael Goin
2026-04-08 18:56:32 -07:00
committed by GitHub
co-authored by Claude
parent 83aea2147f
commit eb4205fee5
12 changed files with 251 additions and 40 deletions
+2 -3
View File
@@ -23,6 +23,7 @@ from vllm.logger import init_logger
from vllm.model_executor.offloader.base import get_offloader
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.utils.deep_gemm import set_num_sms as deep_gemm_set_num_sms
from vllm.utils.import_utils import has_deep_gemm
from vllm.utils.platform_utils import num_compute_units
from vllm.v1.worker.ubatching import UBatchContext, make_ubatch_contexts
@@ -158,9 +159,7 @@ class UBatchWrapper:
# TODO(lucas): support other kernels besides DeepGEMM
set_compute_sms = lambda sms: None
if has_deep_gemm() and comm_sms > 0:
import deep_gemm as dg
set_compute_sms = lambda sms: dg.set_num_sms(sms)
set_compute_sms = lambda sms: deep_gemm_set_num_sms(sms)
return SMControlContextManager(
comm_sms=comm_sms,