A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[V1] [Hybrid] Remove code to override default CUDA graph configuration (#26226)
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Thomas Parnell committed
Oct 4, 2025 at 15:47 UTC
ed3aeb25a4cf833027ce937c5fdfe50371b7fabd
1 file changed
-15
vllm/model_executor/models/config.py
-15
@@ -4,7 +4,6 @@ from copy import deepcopy
4
from typing import TYPE_CHECKING
5
6
import vllm.envs as envs
7
-from vllm.config.compilation import CUDAGraphMode
7
from vllm.logger import init_logger
8
from vllm.model_executor.models import ModelRegistry
9
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, cdiv
@@ -290,7 +289,6 @@ class MambaModelConfig(VerifyAndUpdateConfig):
289
290
model_config = vllm_config.model_config
291
cache_config = vllm_config.cache_config
293
- compilation_config = vllm_config.compilation_config
292
293
# Set mamba block size to max_model_len (this may get
294
# override by prefix caching logic later)
@@ -320,19 +318,6 @@ class MambaModelConfig(VerifyAndUpdateConfig):
318
"for hybrid models.")
319
model_config.disable_cascade_attn = True
320
323
- # TODO(tdoublep): remove as full cuda graph support is added
324
- FCG_NOT_SUPPORTED_MODELS = [
325
- "Lfm2ForCausalLM",
326
- "MiniMaxText01ForCausalLM",
327
- ]
328
-
329
- if (model_config.architecture not in FCG_NOT_SUPPORTED_MODELS
330
- and compilation_config.cudagraph_mode is None):
331
- logger.info(
332
- "Hybrid or mamba-based model detected: setting cudagraph mode "
333
- "to FULL_AND_PIECEWISE in order to optimize performance.")
334
- compilation_config.cudagraph_mode = CUDAGraphMode.FULL_AND_PIECEWISE
335
-
321
322
class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
323