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, 12:15 UTC
add more pytorch related tests for torch nightly (#17422)
Signed-off-by: Yang Wang <elainewy@meta.com>
Yang Wang committed
May 2, 2025 at 03:29 UTC
b8b0859b5c8ab41dd4d233e502bb2640c91d32aa
3 files changed
+14
-4
.buildkite/test-pipeline.yaml
+4
@@ -293,6 +293,7 @@ steps:
293
parallelism: 4
294
295
- label: PyTorch Compilation Unit Tests
296
+ torch_nightly: true
297
source_file_dependencies:
298
- vllm/
299
- tests/compile
@@ -302,6 +303,7 @@ steps:
303
- pytest -v -s compile/test_sequence_parallelism.py
304
305
- label: PyTorch Fullgraph Smoke Test # 9min
306
+ torch_nightly: true
307
source_file_dependencies:
308
- vllm/
309
- tests/compile
@@ -312,6 +314,7 @@ steps:
314
- pytest -v -s compile/piecewise/test_toy_llama.py
315
316
- label: PyTorch Fullgraph Test # 18min
317
+ torch_nightly: true
318
source_file_dependencies:
319
- vllm/
320
- tests/compile
@@ -436,6 +439,7 @@ steps:
439
##### models test #####
440
441
- label: Basic Models Test # 24min
442
+ torch_nightly: true
443
source_file_dependencies:
444
- vllm/
445
- tests/models
requirements/nightly_torch_test.txt
+7
-1
@@ -23,5 +23,11 @@ runai-model-streamer-s3==0.11.0
23
tensorizer>=2.9.0
24
lm-eval==0.4.8
25
buildkite-test-collector==0.1.9
26
-
26
lm-eval[api]==0.4.8 # required for model evaluation test
27
+
28
+# required for quantization test
29
+bitsandbytes>=0.45.3
30
+
31
+# required for minicpmo_26 test
32
+vector_quantize_pytorch
33
+vocos
vllm/sampling_params.py
+3
-3
@@ -186,9 +186,9 @@ class SamplingParams(
186
logits_processors: list of functions that modify logits based on
187
previously generated tokens, and optionally prompt tokens as
188
a first argument.
189
- truncate_prompt_tokens: If set to -1, will use the truncation size
190
- supported by the model. If set to an integer k, will use only
191
- the last k tokens from the prompt (i.e., left truncation).
189
+ truncate_prompt_tokens: If set to -1, will use the truncation size
190
+ supported by the model. If set to an integer k, will use only
191
+ the last k tokens from the prompt (i.e., left truncation).
192
Defaults to None (i.e., no truncation).
193
guided_decoding: If provided, the engine will construct a guided
194
decoding logits processor from these parameters. Defaults to None.