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 15, 17:15 UTC
[Misc] Restrict ray version dependency and update PP feature warning in V1 (#15556)
Rui Qiao committed
Mar 26, 2025 at 23:21 UTC
df8d3d1287c41ea1dfb5847f920ca9e21aafd568
4 files changed
+8
-5
requirements/cuda.txt
+1
-1
@@ -4,7 +4,7 @@
4
numba == 0.60.0 # v0.61 doesn't support Python 3.9. Required for N-gram speculative decoding
5
6
# Dependencies for NVIDIA GPUs
7
-ray[cgraph]>=2.43.0 # Ray Compiled Graph, required for pipeline parallelism in V1.
7
+ray[cgraph]>=2.43.0, !=2.44.* # Ray Compiled Graph, required for pipeline parallelism in V1.
8
torch==2.6.0
9
torchaudio==2.6.0
10
# These must be updated alongside torch
requirements/test.in
+1
-1
@@ -17,7 +17,7 @@ vector_quantize_pytorch # required for minicpmo_26 test
17
vocos # required for minicpmo_26 test
18
peft
19
pqdm
20
-ray[cgraph]>=2.43.0 # Ray Compiled Graph, required by pipeline parallelism tests
20
+ray[cgraph]>=2.43.0, !=2.44.* # Ray Compiled Graph, required by pipeline parallelism tests
21
sentence-transformers # required for embedding tests
22
soundfile # required for audio tests
23
jiwer # required for audio tests
vllm/config.py
+1
-1
@@ -313,7 +313,7 @@ class ModelConfig:
313
raise ValueError(
314
"VLLM_ATTENTION_BACKEND is set to FLASHINFER, but flashinfer "
315
"module was not found."
316
- "See https://github.com/vllm-project/vllm/blob/main/Dockerfile"
316
+ "See https://github.com/vllm-project/vllm/blob/main/Dockerfile "
317
"for instructions on how to install it.")
318
319
# The tokenizer version is consistent with the model version by default.
vllm/engine/arg_utils.py
+5
-2
@@ -1686,8 +1686,11 @@ class EngineArgs:
1686
if self.enable_lora and _warn_or_fallback("LORA"):
1687
return False
1688
1689
- # PP is supported on V1, but off by default for now.
1690
- if self.pipeline_parallel_size > 1 and _warn_or_fallback("PP"):
1689
+ # PP is supported on V1 with Ray distributed executor,
1690
+ # but off for MP distributed executor for now.
1691
+ if (self.pipeline_parallel_size > 1
1692
+ and self.distributed_executor_backend == "mp"
1693
+ and _warn_or_fallback("PP (MP distributed executor)")):
1694
return False
1695
1696
# ngram is supported on V1, but off by default for now.