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
[Bugfix] Added more env vars to hash (#22449)
Signed-off-by: Julien Lin <jullin@nvidia.com>
nvjullin committed
Aug 16, 2025 at 04:08 UTC
79899b63f6d4517e002cc53eb1dcd47cd9e371ea
1 file changed
+36
-10
vllm/envs.py
+36
-10
@@ -1199,14 +1199,6 @@ def compute_hash() -> str:
1199
affect the choice of different kernels or attention backends should
1200
also be included in the factors list.
1201
"""
1202
- factors: list[Any] = []
1203
-
1204
- # summarize environment variables
1205
- def factorize(name: str):
1206
- if __getattr__(name):
1207
- factors.append(__getattr__(name))
1208
- else:
1209
- factors.append("None")
1202
1203
# The values of envs may affects the computation graph.
1204
# TODO(DefTruth): hash all environment variables?
@@ -1221,11 +1213,45 @@ def compute_hash() -> str:
1213
"VLLM_DP_SIZE",
1214
"VLLM_USE_STANDALONE_COMPILE",
1215
"VLLM_FUSED_MOE_CHUNK_SIZE",
1216
+ "VLLM_FLASHINFER_MOE_BACKEND",
1217
+ "VLLM_V1_USE_PREFILL_DECODE_ATTENTION",
1218
+ "VLLM_USE_AITER_UNIFIED_ATTENTION",
1219
+ "VLLM_ATTENTION_BACKEND",
1220
+ "VLLM_USE_FLASHINFER_SAMPLER",
1221
+ "VLLM_FLASHINFER_FORCE_TENSOR_CORES",
1222
+ "VLLM_DISABLED_KERNELS",
1223
+ "VLLM_USE_DEEP_GEMM",
1224
"VLLM_USE_TRTLLM_FP4_GEMM",
1225
+ "VLLM_USE_FLASHINFER_MOE_FP8",
1226
+ "VLLM_USE_FLASHINFER_MOE_FP4",
1227
+ "VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8",
1228
+ "VLLM_USE_FLASHINFER_MOE_MXFP4_BF16",
1229
+ "VLLM_USE_CUDNN_PREFILL",
1230
+ "VLLM_USE_TRTLLM_ATTENTION",
1231
+ "VLLM_ROCM_USE_AITER",
1232
+ "VLLM_ROCM_USE_AITER_PAGED_ATTN",
1233
+ "VLLM_ROCM_USE_AITER_LINEAR",
1234
+ "VLLM_ROCM_USE_AITER_MOE",
1235
+ "VLLM_ROCM_USE_AITER_RMSNORM",
1236
+ "VLLM_ROCM_USE_AITER_MLA",
1237
+ "VLLM_ROCM_USE_AITER_MHA",
1238
+ "VLLM_ROCM_USE_SKINNY_GEMM",
1239
+ "VLLM_ROCM_FP8_PADDING",
1240
+ "VLLM_ROCM_MOE_PADDING",
1241
+ "VLLM_ROCM_CUSTOM_PAGED_ATTN",
1242
+ "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION",
1243
+ "VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16",
1244
+ "VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB",
1245
]
1246
for key in environment_variables_to_hash:
1227
- if key in environment_variables:
1228
- factorize(key)
1247
+ # if this goes out of sync with environment_variables,
1248
+ # it's not a user error, it's a bug
1249
+ assert key in environment_variables, \
1250
+ "Please update environment_variables_to_hash in envs.py"
1251
+
1252
+ factors = [
1253
+ environment_variables[key]() for key in environment_variables_to_hash
1254
+ ]
1255
1256
hash_str = hashlib.md5(str(factors).encode(),
1257
usedforsecurity=False).hexdigest()