Datasets:
Lossfunk
/
ISO-Bench

Dataset card Data Studio Files
xet
 Community
1
commit_hash
string
pr_url
string
pr_date
string
timeline_extracted_at
string
analysis_extracted_at
string
models
list
perf_command
string
has_serving
bool
has_latency
bool
has_throughput
bool
uses_lm_eval
bool
commit_subject
string
commit_message
string
commit_date
string
files_changed
list
stats
dict
diff_text
string
apis
list
affected_paths
list
repo
string
hardware
string
lm_eval_command
string
fc542144c4477ffec1d3de6fa43e54f8fb5351e8
https://github.com/vllm-project/vllm/pull/12563
2025-01-31
2025-09-07 17:46:50
2025-09-07 17:46:50
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --num-prompts 100
true
false
false
true
[Feature] Fix guided decoding blocking bitmask memcpy (#12563)
[Feature] Fix guided decoding blocking bitmask memcpy (#12563) **[Guided decoding performance optimization]** Sending the guided decoding bitmask in xgrammar to the GPU (`self.token_bitmask.to(scores.device)`) is a blocking operation that prevents the CPU from pre-launching the sampler kernels. The CPU waits until dec...
2025-01-31T15:37:30-08:00
[ "vllm/model_executor/guided_decoding/xgrammar_decoding.py" ]
{ "commit_year": 2025, "num_edited_lines": 4, "num_files": 1, "num_hunks": 1, "num_non_test_edited_lines": 4, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/guided_decoding/xgrammar_decoding.py b/vllm/model_executor/guided_decoding/xgrammar_decoding.py index 2d8594cb8..ee30ce96f 100644 --- a/vllm/model_executor/guided_decoding/xgrammar_decoding.py +++ b/vllm/model_executor/guided_decoding/xgrammar_decoding.py @@ -307,8 +307,8 @@ class XGram...
[ "None" ]
[ "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/llm.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
fa63e710c7fbaae3a445f669d3b5ba6b9a4ef412
https://github.com/vllm-project/vllm/pull/12094
2025-01-26
2025-09-07 17:46:54
2025-09-07 17:46:54
[ "meta-llama/Meta-Llama-3-8B" ]
VLLM_USE_V1=1 python benchmarks/benchmark_latency.py --model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --input-len 1000 --batch-size 32
false
true
false
true
[V1][Perf] Reduce scheduling overhead in model runner after cuda sync (#12094)
[V1][Perf] Reduce scheduling overhead in model runner after cuda sync (#12094)
2025-01-26T00:42:37-08:00
[ "vllm/v1/outputs.py", "vllm/v1/sample/sampler.py", "vllm/v1/worker/gpu_model_runner.py" ]
{ "commit_year": 2025, "num_edited_lines": 34, "num_files": 3, "num_hunks": 6, "num_non_test_edited_lines": 34, "num_non_test_files": 3, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/v1/outputs.py b/vllm/v1/outputs.py index acc3a944e..32aee44e3 100644 --- a/vllm/v1/outputs.py +++ b/vllm/v1/outputs.py @@ -8,7 +8,7 @@ import torch class SamplerOutput: # [num_reqs] - sampled_token_ids: List[int] + sampled_token_ids: torch.Tensor # [num_reqs, max_num_logprobs + 1]...
[ "vllm.v1.outputs.SamplerOutput", "vllm.v1.sample.sampler.Sampler.forward", "vllm.v1.worker.GPUModelRunner.execute_model" ]
[ "vllm/v1/worker/gpu_model_runner.py", "vllm/v1/sample/sampler.py", "vllm/model_executor/layers/sampler.py", "vllm/v1/sample/tpu/sampler.py", "vllm/outputs.py", "vllm/v1/outputs.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Meta-Llama-3-8B,dtype=auto --tasks gsm8k --batch_size auto --limit 100
6dd94dbe94c1820a1e224cba65efcf0befa97995
https://github.com/vllm-project/vllm/pull/12380
2025-01-24
2025-09-07 17:46:57
2025-09-07 17:46:57
[ "meta-llama/Meta-Llama-3-8B" ]
python benchmarks/benchmark_latency.py --model meta-llama/Meta-Llama-3-8B --load-format dummy
false
true
false
true
[perf] fix perf regression from #12253 (#12380)
[perf] fix perf regression from #12253 (#12380)
2025-01-24T11:34:27+08:00
[ "vllm/worker/model_runner.py" ]
{ "commit_year": 2025, "num_edited_lines": 5, "num_files": 1, "num_hunks": 2, "num_non_test_edited_lines": 5, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/worker/model_runner.py b/vllm/worker/model_runner.py index cf2f1c6b3..bf1a40d48 100644 --- a/vllm/worker/model_runner.py +++ b/vllm/worker/model_runner.py @@ -455,7 +455,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]): self.enable_prompt_adapter = (self.runne...
[ "vllm.worker.model_runner.ModelInputForGPUBuilder.__init__" ]
[ "vllm/worker/model_runner.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/openai/serving_completion.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Meta-Llama-3-8B,dtype=auto --tasks gsm8k --batch_size auto --limit 100
310aca88c984983189a57f1b72e3b1dde89fb92f
https://github.com/vllm-project/vllm/pull/11870
2025-01-09
2025-09-07 17:47:12
2025-09-07 17:47:12
[ "meta-llama/Meta-Llama-3-70B" ]
python benchmarks/benchmark_latency.py --model meta-llama/Meta-Llama-3-70B --load-format dummy --enforce-eager -tp 4
false
true
false
true
[perf]fix current stream (#11870)
[perf]fix current stream (#11870)
2025-01-09T07:18:21Z
[ "vllm/distributed/device_communicators/pynccl.py", "vllm/distributed/parallel_state.py", "vllm/utils.py", "vllm/worker/multi_step_model_runner.py" ]
{ "commit_year": 2025, "num_edited_lines": 61, "num_files": 4, "num_hunks": 14, "num_non_test_edited_lines": 61, "num_non_test_files": 4, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/distributed/device_communicators/pynccl.py b/vllm/distributed/device_communicators/pynccl.py index fda4d007c..efc599871 100644 --- a/vllm/distributed/device_communicators/pynccl.py +++ b/vllm/distributed/device_communicators/pynccl.py @@ -10,6 +10,7 @@ from vllm.distributed.device_communicators.pynccl...
[ "vllm.distributed.device_communicators.pynccl.PyNcclCommunicator.all_reduce", "vllm.utils.current_stream", "vllm.worker.multi_step_model_runner.MultiStepModelRunner.execute_model" ]
[ "vllm/distributed/device_communicators/pynccl.py", "vllm/distributed/parallel_state.py", "vllm/worker/multi_step_model_runner.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Meta-Llama-3-70B,dtype=auto --tasks gsm8k --batch_size auto --limit 100
b55ed6ef8ab0dce7fb0f79ff292dafdb4d22610c
https://github.com/vllm-project/vllm/pull/11692
2025-01-02
2025-09-07 17:47:18
2025-09-07 17:47:18
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm
true
false
false
true
[V1][Minor] Optimize token_ids_cpu copy (#11692)
[V1][Minor] Optimize token_ids_cpu copy (#11692)
2025-01-02T12:04:58-07:00
[ "vllm/v1/worker/gpu_input_batch.py", "vllm/v1/worker/gpu_model_runner.py" ]
{ "commit_year": 2025, "num_edited_lines": 14, "num_files": 2, "num_hunks": 4, "num_non_test_edited_lines": 14, "num_non_test_files": 2, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/v1/worker/gpu_input_batch.py b/vllm/v1/worker/gpu_input_batch.py index e79145300..f8a1427c6 100644 --- a/vllm/v1/worker/gpu_input_batch.py +++ b/vllm/v1/worker/gpu_input_batch.py @@ -66,8 +66,9 @@ class InputBatch: pin_memory=False, ) self.token_ids_cpu = self.token_ids_...
[ "InputBatch.add_request", "InputBatch.condense", "GPUModelRunner._update_states" ]
[ "vllm/v1/worker/gpu_input_batch.py", "vllm/v1/worker/gpu_model_runner.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
3b61cb450d899dc423feb264c297d4d18d701678
https://github.com/vllm-project/vllm/pull/10989
2024-12-09
2025-09-07 17:47:34
2025-09-07 17:47:34
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_latency.py --model meta-llama/Llama-3.1-8B-Instruct --batch-size 32 --input-len 512 --output-len 128
false
true
false
true
[V1] Further reduce CPU overheads in flash-attn (#10989)
[V1] Further reduce CPU overheads in flash-attn (#10989)
2024-12-09T12:38:46-08:00
[ "csrc/cache_kernels.cu", "vllm/v1/attention/backends/flash_attn.py" ]
{ "commit_year": 2024, "num_edited_lines": 35, "num_files": 2, "num_hunks": 2, "num_non_test_edited_lines": 35, "num_non_test_files": 2, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/csrc/cache_kernels.cu b/csrc/cache_kernels.cu index 1be806bbf..8a95279f9 100644 --- a/csrc/cache_kernels.cu +++ b/csrc/cache_kernels.cu @@ -307,10 +307,20 @@ void reshape_and_cache_flash( torch::Tensor& key_cache, // [num_blocks, block_size, num_heads, head_size] torch::Tensor& value_ca...
[ "vllm.v1.attention.backends.flash_attn.FlashAttentionImpl.forward", "torch.ops._C_cache_ops.reshape_and_cache_flash" ]
[ "vllm/attention/backends/flash_attn.py", "vllm/v1/attention/backends/flash_attn.py", "vllm/_custom_ops.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
98f47f2a4032f8c395268de80858c64ffcfc60fa
https://github.com/vllm-project/vllm/pull/10733
2024-11-28
2025-09-07 17:47:41
2025-09-07 17:47:41
[ "facebook/opt-125m" ]
python benchmarks/benchmark_latency.py --model facebook/opt-125m
false
true
false
true
[V1] Optimize the CPU overheads in FlashAttention custom op (#10733)
[V1] Optimize the CPU overheads in FlashAttention custom op (#10733)
2024-11-28T09:01:02-08:00
[ "vllm/v1/attention/backends/flash_attn.py" ]
{ "commit_year": 2024, "num_edited_lines": 17, "num_files": 1, "num_hunks": 4, "num_non_test_edited_lines": 17, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py index 5f8535eaa..e618edf7d 100644 --- a/vllm/v1/attention/backends/flash_attn.py +++ b/vllm/v1/attention/backends/flash_attn.py @@ -135,6 +135,13 @@ class FlashAttentionImpl(AttentionImpl): assert k_scale == 1.0 an...
[ "vllm.v1.attention.backends.flash_attn.FlashAttentionImpl.forward", "vllm.v1.attention.backends.flash_attn.unified_v1_flash_attention" ]
[ "vllm/attention/backends/flash_attn.py", "vllm/v1/attention/backends/flash_attn.py", "vllm/_custom_ops.py", "csrc/torch_bindings.cpp", "csrc/cpu/torch_bindings.cpp", "csrc/rocm/torch_bindings.cpp", "csrc/moe/torch_bindings.cpp" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=facebook/opt-125m,dtype=auto --tasks gsm8k --batch_size auto --limit 100
8c1e77fb585c4f42783a3d88c1efc7c9e15fd89f
https://github.com/vllm-project/vllm/pull/10742
2024-11-28
2025-09-07 17:47:44
2025-09-07 17:47:44
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_latency.py --model meta-llama/Llama-3.1-8B-Instruct --batch-size 32 --input-len 512 --output-len 128
false
true
false
true
[Kernel] Update vllm-flash-attn version to reduce CPU overheads (#10742)
[Kernel] Update vllm-flash-attn version to reduce CPU overheads (#10742)
2024-11-28T08:31:28-08:00
[ "CMakeLists.txt" ]
{ "commit_year": 2024, "num_edited_lines": 2, "num_files": 1, "num_hunks": 1, "num_non_test_edited_lines": 2, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/CMakeLists.txt b/CMakeLists.txt index 45a3b484e..f43bf8143 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -522,7 +522,7 @@ else() FetchContent_Declare( vllm-flash-attn GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git - GIT_TAG d886f88165702b3c7e77445...
[ "None" ]
[ "vllm/attention/backends/flash_attn.py", "vllm/v1/attention/backends/flash_attn.py", "vllm/attention/ops/triton_flash_attention.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
6e36f4fa6ce64619b9ea94c88a157f5783a63a65
https://github.com/vllm-project/vllm/pull/7874
2024-09-02
2025-09-07 17:48:01
2025-09-07 17:48:01
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
improve chunked prefill performance
improve chunked prefill performance [Bugfix] Fix #7592 vllm 0.5.4 enable_chunked_prefill throughput is slightly lower than 0.5.3~0.5.0. (#7874)
2024-09-02T14:20:12-07:00
[ "tests/basic_correctness/test_chunked_prefill.py", "vllm/core/scheduler.py" ]
{ "commit_year": 2024, "num_edited_lines": 18, "num_files": 2, "num_hunks": 2, "num_non_test_edited_lines": 15, "num_non_test_files": 1, "num_test_files": 1, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/basic_correctness/test_chunked_prefill.py b/tests/basic_correctness/test_chunked_prefill.py index fc6f829c3..a63ac380e 100644 --- a/tests/basic_correctness/test_chunked_prefill.py +++ b/tests/basic_correctness/test_chunked_prefill.py @@ -116,6 +116,9 @@ def test_models_with_fp8_kv_cache( pyt...
[ "vllm.core.scheduler.Scheduler.schedule", "vllm.core.scheduler.SchedulerOutputs" ]
[ "vllm/core/scheduler.py", "vllm/v1/core/sched/scheduler.py", "vllm/attention/ops/chunked_prefill_paged_decode.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
e3580537a41a46b0f3cd750b86b633c1857a8c90
https://github.com/vllm-project/vllm/pull/7753
2024-08-28
2025-09-07 17:48:09
2025-09-07 17:48:09
[ "RedHatAI/Meta-Llama-3-8B-Instruct-FP8" ]
python benchmarks/benchmark_serving.py --model RedHatAI/Meta-Llama-3-8B-Instruct-FP8 --enable-prefix-caching --enable-chunked-prefill --max-num-batched-tokens 2048
true
false
false
true
[Performance] Enable chunked prefill and prefix caching together (#7753)
[Performance] Enable chunked prefill and prefix caching together (#7753)
2024-08-28T00:36:31-07:00
[ "tests/basic_correctness/test_chunked_prefill.py", "tests/core/test_block_manager.py", "tests/core/test_chunked_prefill_scheduler.py", "vllm/core/block_manager_v1.py", "vllm/core/block_manager_v2.py", "vllm/core/embedding_model_block_manager.py", "vllm/core/interfaces.py", "vllm/core/scheduler.py", ...
{ "commit_year": 2024, "num_edited_lines": 252, "num_files": 9, "num_hunks": 12, "num_non_test_edited_lines": 107, "num_non_test_files": 6, "num_test_files": 3, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/basic_correctness/test_chunked_prefill.py b/tests/basic_correctness/test_chunked_prefill.py index 1211e6ba5..fc6f829c3 100644 --- a/tests/basic_correctness/test_chunked_prefill.py +++ b/tests/basic_correctness/test_chunked_prefill.py @@ -6,6 +6,7 @@ prefill requests are chunked. Run `pytest tests/...
[ "ModelRunner.generate_greedy", "Scheduler.schedule", "BlockSpaceManager.mark_blocks_as_computed" ]
[ "vllm/worker/model_runner.py", "vllm/core/scheduler.py", "vllm/v1/core/sched/scheduler.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=RedHatAI/Meta-Llama-3-8B-Instruct-FP8,dtype=auto --tasks gsm8k --batch_size auto --limit 100
2deb029d115dadd012ce5ea70487a207cb025493
https://github.com/vllm-project/vllm/pull/7822
2024-08-26
2025-09-07 17:48:12
2025-09-07 17:48:12
[ "RedHatAI/Meta-Llama-3-8B-Instruct-FP8" ]
python benchmarks/benchmark_prefix_caching.py --model RedHatAI/Meta-Llama-3-8B-Instruct-FP8 --output-len 200 --enable-prefix-caching
true
false
false
true
[Performance][BlockManagerV2] Mark prefix cache block as computed after schedule (#7822)
[Performance][BlockManagerV2] Mark prefix cache block as computed after schedule (#7822)
2024-08-26T11:24:53-07:00
[ "tests/core/block/test_prefix_caching_block.py", "vllm/core/block/prefix_caching_block.py", "vllm/core/block_manager_v2.py" ]
{ "commit_year": 2024, "num_edited_lines": 63, "num_files": 3, "num_hunks": 6, "num_non_test_edited_lines": 32, "num_non_test_files": 2, "num_test_files": 1, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/core/block/test_prefix_caching_block.py b/tests/core/block/test_prefix_caching_block.py index c2226870c..25be2dd13 100644 --- a/tests/core/block/test_prefix_caching_block.py +++ b/tests/core/block/test_prefix_caching_block.py @@ -708,6 +708,37 @@ class TestPrefixCachingBlockAllocator: ...
[ "PrefixCachingBlockAllocator.mark_blocks_as_computed", "BlockSpaceManagerV2.mark_blocks_as_computed" ]
[ "vllm/core/block/prefix_caching_block.py", "vllm/core/block_manager.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=RedHatAI/Meta-Llama-3-8B-Instruct-FP8,dtype=auto --tasks gsm8k --batch_size auto --limit 100
fc7b8d1eefcbe837a56b7c080509417fe5167e6c
https://github.com/vllm-project/vllm/pull/7364
2024-08-09
2025-09-07 17:48:14
2025-09-07 17:48:14
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
[Performance] e2e overheads reduction: Small followup diff (#7364)
[Performance] e2e overheads reduction: Small followup diff (#7364)
2024-08-09T15:49:36Z
[ "vllm/core/block_manager_v1.py", "vllm/sequence.py" ]
{ "commit_year": 2024, "num_edited_lines": 7, "num_files": 2, "num_hunks": 2, "num_non_test_edited_lines": 7, "num_non_test_files": 2, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/core/block_manager_v1.py b/vllm/core/block_manager_v1.py index 622aca66a..ad26d3c51 100644 --- a/vllm/core/block_manager_v1.py +++ b/vllm/core/block_manager_v1.py @@ -336,9 +336,9 @@ class BlockSpaceManagerV1(BlockSpaceManager): # Assign the self-attention block tables for each sequence. ...
[ "BlockSpaceManagerV1", "SequenceGroup.get_finished_seqs" ]
[ "vllm/sequence.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
660470e5a36b8e52083615ad7c85e9b4fd4c72ce
https://github.com/vllm-project/vllm/pull/7193
2024-08-06
2025-09-07 17:48:19
2025-09-07 17:48:19
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --tensor-parallel-size 1 --enable-prefix-caching --use-v2-block-manager
true
false
false
true
[Core] Optimize evictor-v2 performance (#7193)
[Core] Optimize evictor-v2 performance (#7193)
2024-08-06T12:34:25-07:00
[ "vllm/core/evictor_v2.py" ]
{ "commit_year": 2024, "num_edited_lines": 6, "num_files": 1, "num_hunks": 2, "num_non_test_edited_lines": 6, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/core/evictor_v2.py b/vllm/core/evictor_v2.py index 3dd12e2e2..5b1a208b7 100644 --- a/vllm/core/evictor_v2.py +++ b/vllm/core/evictor_v2.py @@ -91,8 +91,9 @@ class LRUEvictor(Evictor): # at the start of OrderedDict. Loop through all these blocks to # find the one with maximum number o...
[ "None" ]
[ "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/api_server.py", "vllm/entrypoints/openai/api_server.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
89a84b0bb7b30706a02836234a94493ea8f780bf
https://github.com/vllm-project/vllm/pull/6779
2024-07-26
2025-09-07 17:48:26
2025-09-07 17:48:26
[ "Qwen/Qwen1.5-0.5B" ]
python benchmarks/benchmark_serving.py --model Qwen/Qwen1.5-0.5B --backend vllm --num-prompts 2048 --input-len 1024
true
false
false
true
[Core] Use array to speedup padding (#6779)
[Core] Use array to speedup padding (#6779)
2024-07-25T21:31:31-07:00
[ "vllm/model_executor/layers/sampler.py", "vllm/model_executor/sampling_metadata.py", "vllm/sequence.py" ]
{ "commit_year": 2024, "num_edited_lines": 46, "num_files": 3, "num_hunks": 9, "num_non_test_edited_lines": 46, "num_non_test_files": 3, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/layers/sampler.py b/vllm/model_executor/layers/sampler.py index 5c376797a..121458f81 100644 --- a/vllm/model_executor/layers/sampler.py +++ b/vllm/model_executor/layers/sampler.py @@ -220,7 +220,7 @@ def _apply_min_tokens_penalty( seqs_to_penalize: List[int] = [] ...
[ "vllm.sequence.SequenceData.prompt_token_ids_array", "vllm.sequence.SequenceData.output_token_ids_array", "vllm.model_executor.sampling_metadata.SamplingTensors.from_lists" ]
[ "vllm/sequence.py", "vllm/model_executor/sampling_metadata.py", "vllm/v1/sample/sampler.py", "vllm/model_executor/layers/sampler.py", "vllm/v1/sample/tpu/sampler.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=Qwen/Qwen1.5-0.5B,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
9ed82e7074a18e25680ab106fc846364ad97bc00
https://github.com/vllm-project/vllm/pull/6520
2024-07-19
2025-09-07 17:48:29
2025-09-07 17:48:29
[ "meta-llama/Llama-2-7b-hf" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-2-7b-hf --backend vllm --num-prompts 100
true
false
false
true
[Misc] Small perf improvements (#6520)
[Misc] Small perf improvements (#6520)
2024-07-19T12:10:56-07:00
[ "tests/core/block/test_block_manager_v2.py", "tests/core/block/test_cpu_gpu_block_allocator.py", "vllm/core/block/block_table.py", "vllm/core/block/prefix_caching_block.py", "vllm/model_executor/models/__init__.py", "vllm/sequence.py", "vllm/utils.py" ]
{ "commit_year": 2024, "num_edited_lines": 69, "num_files": 7, "num_hunks": 11, "num_non_test_edited_lines": 50, "num_non_test_files": 5, "num_test_files": 2, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/core/block/test_block_manager_v2.py b/tests/core/block/test_block_manager_v2.py index d0ca09c4b..d7863a9ae 100644 --- a/tests/core/block/test_block_manager_v2.py +++ b/tests/core/block/test_block_manager_v2.py @@ -249,10 +249,13 @@ def test_append_slots(block_size, prompt_len, num_slots_to_append, ...
[ "BlockTable.get_num_token_blocks", "ModelRegistry.load_model_cls", "SequenceGroup.prompt", "SequenceGroup.prompt_token_ids", "SequenceGroup.multi_modal_data" ]
[]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-2-7b-hf,dtype=auto --tasks gsm8k --batch_size auto --limit 100
3476ed0809ec91a3457da0cb90543133a4f4b519
https://github.com/vllm-project/vllm/pull/5602
2024-07-02
2025-09-07 17:48:40
2025-09-07 17:48:40
[ "meta-llama/Llama-2-7b-hf" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-2-7b-hf --dtype float16 --num-prompts 300 --seed 0
true
false
false
true
[Core] Optimize block_manager_v2 vs block_manager_v1 (to make V2 default) (#5602)
[Core] Optimize block_manager_v2 vs block_manager_v1 (to make V2 default) (#5602)
2024-07-01T20:10:37-07:00
[ "benchmarks/benchmark_latency.py", "tests/conftest.py", "tests/core/block/test_block_table.py", "tests/core/block/test_cpu_gpu_block_allocator.py", "tests/core/block/test_naive_block.py", "tests/core/block/test_prefix_caching_block.py", "tests/spec_decode/test_batch_expansion.py", "vllm/core/block/blo...
{ "commit_year": 2024, "num_edited_lines": 1721, "num_files": 19, "num_hunks": 107, "num_non_test_edited_lines": 1570, "num_non_test_files": 13, "num_test_files": 6, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/benchmarks/benchmark_latency.py b/benchmarks/benchmark_latency.py index a46ee1581..8d0554b0f 100644 --- a/benchmarks/benchmark_latency.py +++ b/benchmarks/benchmark_latency.py @@ -46,6 +46,7 @@ def main(args: argparse.Namespace): load_format=args.load_format, distributed_executor_backend=...
[ "CpuGpuBlockAllocator.allocate_mutable_block", "CpuGpuBlockAllocator.allocate_immutable_block", "PrefixCachingBlockAllocator.allocate_mutable_block", "PrefixCachingBlockAllocator.allocate_immutable_block", "PrefixCachingBlock.__init__" ]
[ "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/openai/serving_completion.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-2-7b-hf,dtype=auto --tasks gsm8k --batch_size auto --limit 100
d7740ea4dcee4ab75d7d6eef723f33cae957b288
https://github.com/vllm-project/vllm/pull/4594
2024-05-08
2025-09-07 17:49:00
2025-09-07 17:49:00
[ "meta-llama/Llama-2-7b-hf" ]
python benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --input-len 256 --output-len 256
false
false
true
true
[Core] Optimize sampler get_logprobs (#4594)
[Core] Optimize sampler get_logprobs (#4594)
2024-05-08T08:42:28-07:00
[ "vllm/model_executor/layers/sampler.py" ]
{ "commit_year": 2024, "num_edited_lines": 117, "num_files": 1, "num_hunks": 3, "num_non_test_edited_lines": 117, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/layers/sampler.py b/vllm/model_executor/layers/sampler.py index 1f19d2053..e52e350d2 100644 --- a/vllm/model_executor/layers/sampler.py +++ b/vllm/model_executor/layers/sampler.py @@ -782,13 +782,14 @@ def _get_logprobs( top_logprobs, top_token_ids = torch.topk(logprobs, ...
[ "None" ]
[ "vllm/v1/sample/sampler.py", "vllm/model_executor/layers/sampler.py", "vllm/v1/sample/tpu/sampler.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/llm.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-2-7b-hf,dtype=auto --tasks gsm8k --batch_size auto --limit 100
ad8d696a99ca1eee19f1404e16e8e82df592ff85
https://github.com/vllm-project/vllm/pull/4270
2024-04-22
2025-09-07 17:49:06
2025-09-07 17:49:06
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
[Core] Scheduler perf fix (#4270)
[Core] Scheduler perf fix (#4270)
2024-04-22T21:11:06Z
[ "tests/core/test_scheduler.py", "vllm/core/scheduler.py" ]
{ "commit_year": 2024, "num_edited_lines": 25, "num_files": 2, "num_hunks": 13, "num_non_test_edited_lines": 7, "num_non_test_files": 1, "num_test_files": 1, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/core/test_scheduler.py b/tests/core/test_scheduler.py index 9588a1bea..a25112385 100644 --- a/tests/core/test_scheduler.py +++ b/tests/core/test_scheduler.py @@ -540,7 +540,7 @@ def test_decode_schedule_preempted(): curr_loras = None for i in range(3): _, seq_group = create_dummy_p...
[ "vllm.core.scheduler.Scheduler._allocate_and_set_running" ]
[ "vllm/core/scheduler.py", "vllm/v1/core/sched/scheduler.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/llm.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
9474e89ba4ecae253b585eb6b3e1d85f4e108f01
https://github.com/vllm-project/vllm/pull/3357
2024-03-20
2025-09-07 17:49:25
2025-09-07 17:49:25
[ "huggyllama/llama-7b" ]
python benchmarks/benchmark_throughput.py --model huggyllama/llama-7b --dataset-name sharegpt --num-prompts 2000
false
false
true
true
[PREFIX CACHING FOLLOW UP] A bunch of fixes to block allocator performance when automatic prefix caching is disabled (#3357)
[PREFIX CACHING FOLLOW UP] A bunch of fixes to block allocator performance when automatic prefix caching is disabled (#3357)
2024-03-20T00:11:11-07:00
[ "tests/core/test_block_manager.py", "tests/prefix_caching/test_prefix_caching.py", "vllm/core/block_manager.py", "vllm/core/evictor.py" ]
{ "commit_year": 2024, "num_edited_lines": 286, "num_files": 4, "num_hunks": 25, "num_non_test_edited_lines": 260, "num_non_test_files": 2, "num_test_files": 2, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/core/test_block_manager.py b/tests/core/test_block_manager.py index 44ac05a14..9473a33f0 100644 --- a/tests/core/test_block_manager.py +++ b/tests/core/test_block_manager.py @@ -4,7 +4,7 @@ from typing import List from vllm import SamplingParams from vllm.block import PhysicalTokenBlock -from vll...
[ "vllm.core.block_manager.UncachedBlockAllocator", "vllm.core.block_manager.CachedBlockAllocator", "vllm.core.block_manager.BlockSpaceManager" ]
[ "vllm/core/block_manager.py", "vllm/core/block/prefix_caching_block.py", "examples/offline_inference/automatic_prefix_caching.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=huggyllama/llama-7b,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
b690e34824fd5a5c4054a0c0468ebfb6aa1dd215
https://github.com/vllm-project/vllm/pull/21075
2025-08-02
2025-09-07 17:49:40
2025-09-07 17:49:40
[ "ibm-ai-platform/Bamba-9B-v2" ]
python benchmarks/benchmark_serving.py --model ibm-ai-platform/Bamba-9B-v2 --dataset-name sharegpt
true
false
false
true
[Model] Mamba2 preallocate SSM output tensor to avoid d2d copy overhead (#21075)
[Model] Mamba2 preallocate SSM output tensor to avoid d2d copy overhead (#21075)
2025-08-02T01:59:34-07:00
[ "tests/kernels/mamba/test_mamba_ssm.py", "tests/kernels/mamba/test_mamba_ssm_ssd.py", "vllm/model_executor/layers/mamba/mamba_mixer.py", "vllm/model_executor/layers/mamba/mamba_mixer2.py", "vllm/model_executor/layers/mamba/ops/mamba_ssm.py", "vllm/model_executor/layers/mamba/ops/ssd_chunk_scan.py", "vll...
{ "commit_year": 2025, "num_edited_lines": 262, "num_files": 9, "num_hunks": 39, "num_non_test_edited_lines": 165, "num_non_test_files": 7, "num_test_files": 2, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/kernels/mamba/test_mamba_ssm.py b/tests/kernels/mamba/test_mamba_ssm.py index 8dece26dd..4c32ae81b 100644 --- a/tests/kernels/mamba/test_mamba_ssm.py +++ b/tests/kernels/mamba/test_mamba_ssm.py @@ -365,6 +365,7 @@ def test_selective_state_update(dim, dstate, has_z, itype): batch_size = 1 st...
[ "vllm.model_executor.layers.mamba.ops.mamba_ssm.selective_state_update", "vllm.model_executor.layers.mamba.ops.ssd_combined.mamba_chunk_scan_combined" ]
[ "vllm/model_executor/layers/mamba/mamba_mixer2.py", "vllm/model_executor/models/phi4flash.py", "vllm/model_executor/models/plamo2.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=ibm-ai-platform/Bamba-9B-v2,dtype=float16 --tasks gsm8k --batch_size auto --limit 100
58eee5f2e05b74eb2cb1a3bbda9c04df4805e4cc
https://github.com/vllm-project/vllm/pull/20000
2025-08-02
2025-09-07 17:49:44
2025-09-07 17:49:44
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
[PERF] Use faster way of decode in tokenizer: avoid useless list-to-list conversion (#20000)
[PERF] Use faster way of decode in tokenizer: avoid useless list-to-list conversion (#20000)
2025-08-02T01:43:52-07:00
[ "vllm/transformers_utils/tokenizer.py" ]
{ "commit_year": 2025, "num_edited_lines": 7, "num_files": 1, "num_hunks": 1, "num_non_test_edited_lines": 7, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/transformers_utils/tokenizer.py b/vllm/transformers_utils/tokenizer.py index 24ddd35ab..6a31a4198 100644 --- a/vllm/transformers_utils/tokenizer.py +++ b/vllm/transformers_utils/tokenizer.py @@ -50,11 +50,12 @@ def decode_tokens( `skip_special_tokens=None` means to use the backend's default ...
[ "vllm.transformers_utils.tokenizer.decode_tokens" ]
[ "vllm/transformers_utils/tokenizer.py", "vllm/transformers_utils/tokenizer_base.py", "vllm/transformers_utils/tokenizer_group.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
a32237665df876fcb51196dc209e8aff9fd89d29
https://github.com/vllm-project/vllm/pull/21245
2025-07-22
2025-09-07 17:50:16
2025-09-07 17:50:16
[ "facebook/opt-125m" ]
vllm bench serve --dataset-name random --model facebook/opt-125m --served-model-name facebook/opt-125m --random-input-len 700 --random-output-len 1 --endpoint /v1/completions --ignore-eos --host localhost --port 8000 --request-rate 200 --num-prompts 100
true
false
false
true
[Core] Optimize update checks in LogitsProcessor (#21245)
[Core] Optimize update checks in LogitsProcessor (#21245)
2025-07-22T05:27:18-07:00
[ "vllm/v1/sample/logits_processor.py" ]
{ "commit_year": 2025, "num_edited_lines": 18, "num_files": 1, "num_hunks": 3, "num_non_test_edited_lines": 18, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/v1/sample/logits_processor.py b/vllm/v1/sample/logits_processor.py index 3a4c25964..3a06e7105 100644 --- a/vllm/v1/sample/logits_processor.py +++ b/vllm/v1/sample/logits_processor.py @@ -335,14 +335,19 @@ class LogitBiasLogitsProcessor(LogitsProcessor): if not batch_update: retur...
[ "LogitBiasLogitsProcessor.update_state", "MinTokensLogitsProcessor.update_state" ]
[ "vllm/v1/sample/logits_processor.py", "vllm/model_executor/layers/logits_processor.py", "vllm/entrypoints/openai/logits_processors.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=facebook/opt-125m,dtype=auto --tasks gsm8k --batch_size auto --limit 100
e7b204268132cb775c139574c1ff4ad7e15c8f66
https://github.com/vllm-project/vllm/pull/21334
2025-07-22
2025-09-07 17:50:20
2025-09-07 17:50:20
[ "meta-llama/Llama-4-Maverick-17B-128E-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-4-Maverick-17B-128E-Instruct --dataset-name sharegpt --num-prompts 100
true
false
false
true
Revert "[Performance] Performance improvements in non-blockwise fp8 CUTLASS MoE (#20762) (#21334)
Revert "[Performance] Performance improvements in non-blockwise fp8 CUTLASS MoE (#20762) (#21334)
2025-07-21T21:49:01-07:00
[ "benchmarks/kernels/benchmark_grouped_gemm_cutlass.py", "csrc/moe/moe_permute_unpermute_op.cu", "tests/kernels/moe/test_cutlass_moe.py", "tests/kernels/moe/test_pplx_cutlass_moe.py", "vllm/model_executor/layers/fused_moe/cutlass_moe.py", "vllm/model_executor/layers/quantization/compressed_tensors/compress...
{ "commit_year": 2025, "num_edited_lines": 212, "num_files": 6, "num_hunks": 30, "num_non_test_edited_lines": 176, "num_non_test_files": 4, "num_test_files": 2, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py b/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py index a6b42406b..1d4e730f9 100644 --- a/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py +++ b/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py @@ -80,11 +80,6 @@ def bench_run( a, s...
[ "vllm.cutlass_moe_fp8", "CutlassExpertsFp8.apply", "CompressedTensorsW8A8Fp8MoECutlassMethod.select_gemm_impl" ]
[ "vllm/model_executor/layers/fused_moe/cutlass_moe.py", "vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-4-Maverick-17B-128E-Instruct,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
015069b01741e9ecb9e604c7fe87fbdfc306ebe5
https://github.com/vllm-project/vllm/pull/17515
2025-05-01
2025-09-07 17:51:10
2025-09-07 17:51:10
[ "Qwen/Qwen3-7B-Instruct" ]
python benchmarks/benchmark_serving.py --model Qwen/Qwen3-7B-Instruct --dataset-name sharegpt --request-rate 1
true
false
false
true
[Misc] Optimize the Qwen3_ReasoningParser extract_reasoning_content (#17515)
[Misc] Optimize the Qwen3_ReasoningParser extract_reasoning_content (#17515)
2025-05-01T03:29:01-07:00
[ "vllm/reasoning/qwen3_reasoning_parser.py" ]
{ "commit_year": 2025, "num_edited_lines": 53, "num_files": 1, "num_hunks": 3, "num_non_test_edited_lines": 53, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/reasoning/qwen3_reasoning_parser.py b/vllm/reasoning/qwen3_reasoning_parser.py index f588f4016..7095034b1 100644 --- a/vllm/reasoning/qwen3_reasoning_parser.py +++ b/vllm/reasoning/qwen3_reasoning_parser.py @@ -1,6 +1,5 @@ # SPDX-License-Identifier: Apache-2.0 -import re from collections.abc impor...
[ "Qwen3ReasoningParser.extract_reasoning_content" ]
[ "vllm/reasoning/qwen3_reasoning_parser.py", "examples/online_serving/openai_chat_completion_with_reasoning.py", "examples/online_serving/openai_chat_completion_with_reasoning_streaming.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=Qwen/Qwen3-7B-Instruct,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
bc7c4d206bbfb56b06d218b6c2971e8ca191db36
https://github.com/vllm-project/vllm/pull/13305
2025-04-23
2025-09-07 17:51:14
2025-09-07 17:51:14
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --dtype float16 --num-prompts 300 --seed 0
true
false
false
true
[Kernel][ROCM] Upstream prefix prefill speed up for vLLM V1 (#13305)
[Kernel][ROCM] Upstream prefix prefill speed up for vLLM V1 (#13305)
2025-04-22T19:11:56-07:00
[ "tests/core/block/e2e/test_correctness.py", "vllm/attention/ops/prefix_prefill.py" ]
{ "commit_year": 2025, "num_edited_lines": 1640, "num_files": 2, "num_hunks": 4, "num_non_test_edited_lines": 1634, "num_non_test_files": 1, "num_test_files": 1, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/core/block/e2e/test_correctness.py b/tests/core/block/e2e/test_correctness.py index e9b537ed5..9e8e315d8 100644 --- a/tests/core/block/e2e/test_correctness.py +++ b/tests/core/block/e2e/test_correctness.py @@ -195,15 +195,15 @@ def test_lookahead_greedy_equality_with_preemption(baseline_llm_generator...
[ "vllm.attention.ops.prefix_prefill.context_attention_fwd" ]
[ "vllm/attention/ops/prefix_prefill.py", "vllm/attention/layer.py", "vllm/model_executor/layers/fused_moe/layer.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
299ebb62b269ce167eb1c71b5e39a1dc1f65ce1c
https://github.com/vllm-project/vllm/pull/16436
2025-04-21
2025-09-07 17:51:17
2025-09-07 17:51:17
[ "Qwen/Qwen2.5-1.5B-Instruct" ]
vllm bench serve --model Qwen/Qwen2.5-1.5B-Instruct --request-rate 1 --num-prompts 100 --random-input-len 1000 --random-output-len 100 --tokenizer Qwen/Qwen2.5-1.5B-Instruct --ignore-eos
true
false
false
true
[Core] Speed up decode by remove synchronizing operation in sampler (#16436)
[Core] Speed up decode by remove synchronizing operation in sampler (#16436)
2025-04-21T18:18:22Z
[ "vllm/model_executor/layers/utils.py" ]
{ "commit_year": 2025, "num_edited_lines": 13, "num_files": 1, "num_hunks": 1, "num_non_test_edited_lines": 13, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/layers/utils.py b/vllm/model_executor/layers/utils.py index a9ef97391..5e56be061 100644 --- a/vllm/model_executor/layers/utils.py +++ b/vllm/model_executor/layers/utils.py @@ -47,10 +47,15 @@ def apply_penalties(logits: torch.Tensor, prompt_tokens_tensor: torch.Tensor, output_t...
[ "vllm.model_executor.layers.utils.apply_penalties" ]
[ "vllm/v1/sample/sampler.py", "vllm/model_executor/layers/sampler.py", "vllm/v1/sample/tpu/sampler.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/openai/serving_completion.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=Qwen/Qwen2.5-1.5B-Instruct,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
35fad35a485eac9195c510731ba4a9d297dfd963
https://github.com/vllm-project/vllm/pull/15478
2025-03-26
2025-09-07 17:51:35
2025-09-07 17:51:35
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm
true
false
false
true
[V1][Sampler] Faster top-k only implementation (#15478)
[V1][Sampler] Faster top-k only implementation (#15478)
2025-03-26T10:56:47-07:00
[ "tests/v1/sample/test_topk_topp_sampler.py", "vllm/v1/sample/ops/topk_topp_sampler.py", "vllm/v1/sample/sampler.py" ]
{ "commit_year": 2025, "num_edited_lines": 96, "num_files": 3, "num_hunks": 7, "num_non_test_edited_lines": 59, "num_non_test_files": 2, "num_test_files": 1, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/v1/sample/test_topk_topp_sampler.py b/tests/v1/sample/test_topk_topp_sampler.py new file mode 100644 index 000000000..8a5076412 --- /dev/null +++ b/tests/v1/sample/test_topk_topp_sampler.py @@ -0,0 +1,37 @@ +# SPDX-License-Identifier: Apache-2.0 +import torch +from torch import Generator + +from vllm...
[ "vllm.v1.sample.ops.topk_topp_sampler.apply_top_k_top_p", "vllm.v1.sample.ops.topk_topp_sampler.apply_top_k_only", "vllm.v1.sample.ops.topk_topp_sampler.TopKTopPSampler.forward_native", "vllm.v1.sample.sampler.Sampler.forward" ]
[ "vllm/v1/sample/ops/topk_topp_sampler.py", "vllm/v1/sample/sampler.py", "vllm/model_executor/layers/sampler.py", "vllm/v1/sample/tpu/sampler.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
296f927f2493908984707354e3cc5d7b2e41650b
https://github.com/vllm-project/vllm/pull/14857
2025-03-21
2025-09-07 17:51:42
2025-09-07 17:51:42
[ "ibm-ai-platform/Bamba-9B" ]
python benchmarks/benchmark_serving.py --model ibm-ai-platform/Bamba-9B --dataset-name sharegpt
true
false
false
true
[Model] RE: Mamba2 Prefill Performance Tweaks: Fixing Flurry of Unnecessary Memory Copies (#14857)
[Model] RE: Mamba2 Prefill Performance Tweaks: Fixing Flurry of Unnecessary Memory Copies (#14857)
2025-03-20T19:21:08-07:00
[ "vllm/model_executor/layers/mamba/mamba_mixer2.py" ]
{ "commit_year": 2025, "num_edited_lines": 13, "num_files": 1, "num_hunks": 2, "num_non_test_edited_lines": 13, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/layers/mamba/mamba_mixer2.py b/vllm/model_executor/layers/mamba/mamba_mixer2.py index fec6d6112..d7a45bc51 100644 --- a/vllm/model_executor/layers/mamba/mamba_mixer2.py +++ b/vllm/model_executor/layers/mamba/mamba_mixer2.py @@ -470,10 +470,11 @@ class MambaMixer2(CustomOp): if ...
[ "MambaMixer2.forward_cuda", "mamba_mixer2" ]
[ "vllm/model_executor/layers/mamba/mamba_mixer2.py", "vllm/model_executor/models/mamba2.py", "vllm/model_executor/models/mamba_cache.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=ibm-ai-platform/Bamba-9B,dtype=float16 --tasks gsm8k --batch_size auto --limit 100
22d33baca2c0c639cfd45c48e99803e56c3efa74
https://github.com/vllm-project/vllm/pull/15150
2025-03-19
2025-09-07 17:51:45
2025-09-07 17:51:45
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
[FrontEnd][Perf] `merge_async_iterators` fast-path for single-prompt requests (#15150)
[FrontEnd][Perf] `merge_async_iterators` fast-path for single-prompt requests (#15150)
2025-03-19T21:04:41Z
[ "vllm/utils.py" ]
{ "commit_year": 2025, "num_edited_lines": 5, "num_files": 1, "num_hunks": 1, "num_non_test_edited_lines": 5, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/utils.py b/vllm/utils.py index 79787303a..9bc081890 100644 --- a/vllm/utils.py +++ b/vllm/utils.py @@ -411,6 +411,11 @@ async def merge_async_iterators( When it yields, it yields a tuple (i, item) where i is the index of the iterator that yields the item. """ + if len(iterators) == 1...
[ "vllm.utils.merge_async_iterators" ]
[ "vllm/entrypoints/api_server.py", "vllm/entrypoints/openai/api_server.py", "vllm/entrypoints/openai/serving_completion.py", "vllm/engine/async_llm_engine.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
99abb8b650c66664cdc84d815b7f306f33bd9881
https://github.com/vllm-project/vllm/pull/14930
2025-03-18
2025-09-07 17:51:49
2025-09-07 17:51:49
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --speculative-model '[ngram]' --num-speculative-tokens 3 --dataset-name sharegpt
true
false
false
true
[V1][Spec Decode] Optimize Rejection Sampler with Triton Kernels (#14930)
[V1][Spec Decode] Optimize Rejection Sampler with Triton Kernels (#14930)
2025-03-18T14:31:54-07:00
[ "tests/v1/sample/test_rejection_sampler.py", "vllm/envs.py", "vllm/v1/outputs.py", "vllm/v1/sample/ops/utils.py", "vllm/v1/sample/rejection_sampler.py", "vllm/v1/spec_decode/metadata.py", "vllm/v1/spec_decode/utils.py", "vllm/v1/worker/gpu_model_runner.py" ]
{ "commit_year": 2025, "num_edited_lines": 1329, "num_files": 8, "num_hunks": 34, "num_non_test_edited_lines": 1098, "num_non_test_files": 7, "num_test_files": 1, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/v1/sample/test_rejection_sampler.py b/tests/v1/sample/test_rejection_sampler.py index 84139a40b..8c423e367 100644 --- a/tests/v1/sample/test_rejection_sampler.py +++ b/tests/v1/sample/test_rejection_sampler.py @@ -6,20 +6,23 @@ import torch import torch.nn.functional as F from vllm.v1.sample.meta...
[ "RejectionSampler.forward", "RejectionSampler.parse_output", "SpecDecodeMetadata.make_dummy", "compiled_softmax" ]
[ "vllm/v1/sample/rejection_sampler.py", "vllm/v1/spec_decode/metadata.py", "vllm/v1/sample/metadata.py", "vllm/v1/pool/metadata.py", "vllm/v1/sample/tpu/metadata.py", "vllm/v1/worker/gpu_model_runner.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct --tasks gsm8k --num_fewshot 5 --limit 100
fe66b34728e5d383e3d19aefc544eeee808c99fb
https://github.com/vllm-project/vllm/pull/14778
2025-03-14
2025-09-07 17:51:55
2025-09-07 17:51:55
[ "ibm-ai-platform/Bamba-9B" ]
python benchmarks/benchmark_serving.py --model ibm-ai-platform/Bamba-9B --dtype float16 --num-prompts 300 --seed 0
true
false
false
true
[Model] Mamba2 Prefill Performance Tweaks: Fixing Flurry of Unnecessary Memory Copies (#14778)
[Model] Mamba2 Prefill Performance Tweaks: Fixing Flurry of Unnecessary Memory Copies (#14778)
2025-03-14T16:36:18-04:00
[ "vllm/model_executor/layers/mamba/mamba_mixer2.py" ]
{ "commit_year": 2025, "num_edited_lines": 30, "num_files": 1, "num_hunks": 2, "num_non_test_edited_lines": 30, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/layers/mamba/mamba_mixer2.py b/vllm/model_executor/layers/mamba/mamba_mixer2.py index b53a540ed..5b19e3f35 100644 --- a/vllm/model_executor/layers/mamba/mamba_mixer2.py +++ b/vllm/model_executor/layers/mamba/mamba_mixer2.py @@ -466,10 +466,17 @@ class MambaMixer2(CustomOp): if ...
[ "MambaMixer2.forward_cuda", "mamba_mixer2" ]
[ "vllm/model_executor/layers/mamba/mamba_mixer2.py", "vllm/model_executor/models/mamba2.py", "vllm/model_executor/models/mamba_cache.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=ibm-ai-platform/Bamba-9B,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
70b808fe1a63322bc6bf5f46a91981a8f6b8af00
https://github.com/vllm-project/vllm/pull/14377
2025-03-11
2025-09-07 17:51:59
2025-09-07 17:51:59
[ "Qwen/Qwen2-VL-7B" ]
python benchmarks/benchmark_serving.py --model Qwen/Qwen2-VL-7B --dataset-name random --request-rate 1
true
false
false
true
[Perf]:Optimize qwen2-vl to reduce cudaMemcpyAsync (#14377)
[Perf]:Optimize qwen2-vl to reduce cudaMemcpyAsync (#14377)
2025-03-11T07:39:56Z
[ "vllm/model_executor/models/qwen2_5_vl.py", "vllm/model_executor/models/qwen2_vl.py" ]
{ "commit_year": 2025, "num_edited_lines": 94, "num_files": 2, "num_hunks": 12, "num_non_test_edited_lines": 94, "num_non_test_files": 2, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py index ef3d28c80..ae48c7794 100644 --- a/vllm/model_executor/models/qwen2_5_vl.py +++ b/vllm/model_executor/models/qwen2_5_vl.py @@ -255,10 +255,12 @@ class Qwen2_5_VisionAttention(nn.Module): return q, k, v ...
[ "Qwen2VisionAttention.forward", "Qwen2VisionBlock.forward", "Qwen2VisionTransformer.forward", "Qwen2_5_VisionAttention.forward", "Qwen2_5_VisionTransformer.forward" ]
[ "vllm/model_executor/models/qwen2_vl.py", "vllm/model_executor/models/qwen2_5_vl.py", "vllm/multimodal/registry.py", "vllm/inputs/registry.py", "vllm/model_executor/models/registry.py", "vllm/transformers_utils/chat_templates/registry.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=Qwen/Qwen2-VL-7B,dtype=auto --trust_remote_code --tasks gsm8k --batch_size auto --limit 100
9f1710f1ace3535920c0bb6d4cc329c36289080e
https://github.com/vllm-project/vllm/pull/13897
2025-03-06
2025-09-07 17:52:14
2025-09-07 17:52:14
[ "deepseek-ai/DeepSeek-V2-Lite-Chat" ]
python benchmarks/benchmark_serving.py --model deepseek-ai/DeepSeek-V2-Lite-Chat --input-len 28000 --output-len 64
true
false
false
true
Fix mla prefill context performance (#13897)
Fix mla prefill context performance (#13897)
2025-03-06T09:35:49-08:00
[ "vllm/attention/backends/mla/common.py", "vllm/v1/attention/backends/mla/common.py" ]
{ "commit_year": 2025, "num_edited_lines": 4, "num_files": 2, "num_hunks": 2, "num_non_test_edited_lines": 4, "num_non_test_files": 2, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/attention/backends/mla/common.py b/vllm/attention/backends/mla/common.py index 8184b0732..109e8496f 100644 --- a/vllm/attention/backends/mla/common.py +++ b/vllm/attention/backends/mla/common.py @@ -1308,7 +1308,7 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]): ) k...
[ "MLACommonImpl.forward" ]
[ "vllm/core/block/common.py", "vllm/attention/backends/mla/common.py", "vllm/v1/attention/backends/mla/common.py", "vllm/attention/ops/prefix_prefill.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,dtype=float16 --tasks gsm8k --batch_size auto --limit 100
9badee53decb3d432dc805336abfb0eb81dfb48f
https://github.com/vllm-project/vllm/pull/14223
2025-03-04
2025-09-07 17:52:18
2025-09-07 17:52:18
[ "meta-llama/Llama-3.2-1B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.2-1B-Instruct --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json
true
false
false
true
Fix performance when `--generation-config` is not `None` (#14223)
Fix performance when `--generation-config` is not `None` (#14223)
2025-03-04T20:59:22+01:00
[ "vllm/entrypoints/llm.py", "vllm/entrypoints/openai/serving_chat.py", "vllm/entrypoints/openai/serving_completion.py", "vllm/entrypoints/openai/serving_transcription.py" ]
{ "commit_year": 2025, "num_edited_lines": 48, "num_files": 4, "num_hunks": 8, "num_non_test_edited_lines": 48, "num_non_test_files": 4, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/entrypoints/llm.py b/vllm/entrypoints/llm.py index 122e2ed86..fc585ee9e 100644 --- a/vllm/entrypoints/llm.py +++ b/vllm/entrypoints/llm.py @@ -244,6 +244,7 @@ class LLM: engine_args, usage_context=UsageContext.LLM_CLASS) self.request_counter = Counter() + self.default_sa...
[ "vLLM.LLM.get_default_sampling_params", "OpenAIServingChat.create_chat_completion", "OpenAIServingCompletion.create_completion", "OpenAIServingTranscription._preprocess_transcription" ]
[ "vllm/entrypoints/llm.py", "vllm/entrypoints/openai/serving_chat.py", "vllm/entrypoints/openai/serving_completion.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.2-1B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
19d98e0c7db96713f0e2201649159431177a56e2
https://github.com/vllm-project/vllm/pull/13625
2025-03-03
2025-09-07 17:52:21
2025-09-07 17:52:21
[ "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct" ]
python benchmarks/benchmark_serving.py --model deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct --dataset-name sharegpt --num-prompts 100
true
false
false
true
[Kernel] Optimize moe intermediate_cache usage (#13625)
[Kernel] Optimize moe intermediate_cache usage (#13625)
2025-03-03T16:29:53-05:00
[ "vllm/model_executor/layers/fused_moe/fused_moe.py" ]
{ "commit_year": 2025, "num_edited_lines": 17, "num_files": 1, "num_hunks": 1, "num_non_test_edited_lines": 17, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py index 00260313e..5336b3c10 100644 --- a/vllm/model_executor/layers/fused_moe/fused_moe.py +++ b/vllm/model_executor/layers/fused_moe/fused_moe.py @@ -1240,15 +1240,20 @@ def fused_experts_impl(hidden_state...
[ "torch.ops.vllm.inplace_fused_experts", "torch.ops.vllm.outplace_fused_experts" ]
[ "vllm/model_executor/layers/fused_moe/fused_moe.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/llm.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct,dtype=float16 --tasks gsm8k --batch_size auto
e206b5433109d298e53451015465b2bf8f03ef0a
https://github.com/vllm-project/vllm/pull/13837
2025-02-26
2025-09-07 17:52:25
2025-09-07 17:52:25
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100 --guided-decoding-backend xgrammar
true
false
false
true
[v0][Core] Use xgrammar shared context to avoid copy overhead for offline engine (#13837)
[v0][Core] Use xgrammar shared context to avoid copy overhead for offline engine (#13837)
2025-02-26T14:58:24+08:00
[ "vllm/model_executor/guided_decoding/xgrammar_decoding.py" ]
{ "commit_year": 2025, "num_edited_lines": 26, "num_files": 1, "num_hunks": 2, "num_non_test_edited_lines": 26, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/model_executor/guided_decoding/xgrammar_decoding.py b/vllm/model_executor/guided_decoding/xgrammar_decoding.py index 329b03a57..e6ba7f5ec 100644 --- a/vllm/model_executor/guided_decoding/xgrammar_decoding.py +++ b/vllm/model_executor/guided_decoding/xgrammar_decoding.py @@ -3,7 +3,6 @@ # noqa: UP007 ...
[ "vllm.model_executor.guided_decoding.xgrammar_decoding.XGrammarLogitsProcessor.clone" ]
[ "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py", "vllm/entrypoints/llm.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
6a417b8600d4d1e57698a91b71a38446e8fc5c45
https://github.com/vllm-project/vllm/pull/13589
2025-02-20
2025-09-07 17:52:28
2025-09-07 17:52:28
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
fix neuron performance issue (#13589)
fix neuron performance issue (#13589)
2025-02-20T10:59:36-08:00
[ "vllm/worker/neuron_worker.py" ]
{ "commit_year": 2025, "num_edited_lines": 4, "num_files": 1, "num_hunks": 2, "num_non_test_edited_lines": 4, "num_non_test_files": 1, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/vllm/worker/neuron_worker.py b/vllm/worker/neuron_worker.py index 5f0eb0019..95e7acd02 100644 --- a/vllm/worker/neuron_worker.py +++ b/vllm/worker/neuron_worker.py @@ -76,7 +76,7 @@ class NeuronWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase): # Set the number of GPU blocks to be t...
[ "NeuronWorker.determine_num_available_blocks", "NeuronWorker.initialize_cache" ]
[ "vllm/worker/neuron_worker.py", "vllm/worker/neuron_model_runner.py", "vllm/platforms/neuron.py", "examples/offline_inference/neuron.py", "vllm/model_executor/model_loader/neuron.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
4c822298981a8f7521492075ff72659985fc4c3f
https://github.com/vllm-project/vllm/pull/13365
2025-02-18
2025-09-07 17:52:34
2025-09-07 17:52:34
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_latency.py --model meta-llama/Llama-3.1-8B-Instruct --speculative-model meta-llama/Llama-3.2-1B-Instruct --num-speculative-tokens 5
false
true
false
true
[V1][Spec Decode] Optimize N-gram matching with Numba (#13365)
[V1][Spec Decode] Optimize N-gram matching with Numba (#13365)
2025-02-18T13:19:58-08:00
[ "requirements-common.txt", "vllm/v1/spec_decode/ngram_proposer.py", "vllm/v1/worker/gpu_model_runner.py" ]
{ "commit_year": 2025, "num_edited_lines": 127, "num_files": 3, "num_hunks": 5, "num_non_test_edited_lines": 127, "num_non_test_files": 3, "num_test_files": 0, "only_non_test_files": 1, "only_test_files": 0 }
diff --git a/requirements-common.txt b/requirements-common.txt index b7c94cbdb..c52980bc7 100644 --- a/requirements-common.txt +++ b/requirements-common.txt @@ -1,6 +1,7 @@ psutil sentencepiece # Required for LLaMA tokenizer. numpy < 2.0.0 +numba == 0.60.0 # v0.61 doesn't support Python 3.9. Required for N-gram spe...
[ "vllm.v1.spec_decode.ngram_proposer.NgramProposer.propose", "vllm.v1.worker.GPUModelRunner.__init__" ]
[ "vllm/v1/spec_decode/ngram_proposer.py", "vllm/v1/worker/gpu_model_runner.py", "vllm/engine/llm_engine.py", "vllm/v1/engine/llm_engine.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100
30172b4947c52890b808c6da3a6c7580f55cbb74
https://github.com/vllm-project/vllm/pull/13244
2025-02-18
2025-09-07 17:52:38
2025-09-07 17:52:38
[ "meta-llama/Llama-3.1-8B-Instruct" ]
python benchmarks/benchmark_serving.py --model meta-llama/Llama-3.1-8B-Instruct --backend vllm --num-prompts 100
true
false
false
true
[V1] Optimize handling of sampling metadata and req_ids list (#13244)
[V1] Optimize handling of sampling metadata and req_ids list (#13244)
2025-02-18T12:15:33-08:00
[ "tests/v1/sample/test_rejection_sampler.py", "tests/v1/sample/test_sampler.py", "tests/v1/worker/test_gpu_input_batch.py", "tests/v1/worker/test_gpu_model_runner.py", "vllm/model_executor/layers/utils.py", "vllm/v1/core/scheduler.py", "vllm/v1/sample/metadata.py", "vllm/v1/sample/ops/penalties.py", ...
{ "commit_year": 2025, "num_edited_lines": 553, "num_files": 15, "num_hunks": 74, "num_non_test_edited_lines": 420, "num_non_test_files": 11, "num_test_files": 4, "only_non_test_files": 0, "only_test_files": 0 }
diff --git a/tests/v1/sample/test_rejection_sampler.py b/tests/v1/sample/test_rejection_sampler.py index 8bc33e841..3e810e525 100644 --- a/tests/v1/sample/test_rejection_sampler.py +++ b/tests/v1/sample/test_rejection_sampler.py @@ -26,17 +26,13 @@ def create_logits_tensor(token_ids: List[int], def create_sampling_met...
[ "vllm.v1.sample.metadata.SamplingMetadata", "vllm.v1.sample.sampler.Sampler", "vllm.v1.worker.gpu_input_batch.InputBatch", "vllm.v1.worker.gpu_model_runner.GPUModelRunner" ]
[ "vllm/v1/spec_decode/metadata.py", "vllm/v1/sample/metadata.py", "vllm/v1/pool/metadata.py", "vllm/v1/sample/tpu/metadata.py", "vllm/v1/sample/sampler.py", "vllm/model_executor/layers/sampler.py", "vllm/v1/sample/tpu/sampler.py", "vllm/v1/worker/gpu_input_batch.py" ]
vllm
H100
lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,dtype=auto --tasks gsm8k --batch_size auto --limit 100