Ryan Nguyen fc542144c4 [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 decode is complete, then copies the bitmask over. This PR changes
the operation to async via setting `non-blocking=True`.

(Current) The CPU is blocked on a `cudaStreamSynchronize` and only
pre-empts the sampling kernels after bitmask application. Below is the
Nsys profile for one decode phase from Llama 3.1 8B.

![image](https://github.com/user-attachments/assets/8997eae1-b822-4f52-beb8-ef19a7c6b824)

With the optimization, this is no longer the case:

![image](https://github.com/user-attachments/assets/6d5ea83f-f169-4f98-a8c1-41c719b3e1e7)

---------

Signed-off-by: Ryan N <ryan.nguyen@centml.ai>

2025-01-31 15:37:30 -08:00

..

adapter_commons

[Misc] Clean up and consolidate LRUCache (#11339)

2024-12-20 00:59:32 +08:00

assets

Set weights_only=True when using torch.load() (#12366)

2025-01-24 02:17:30 +00:00

attention

[Attention] MLA decode optimizations (#12528)

2025-01-30 23:49:37 -08:00

compilation

[torch.compile] decouple compile sizes and cudagraph sizes (#12243)

2025-01-24 02:01:30 +08:00

core

Update pre-commit hooks (#12475)

2025-01-27 17:23:08 -07:00

device_allocator

[Core] Support fully transparent sleep mode (#11743)

2025-01-22 14:39:32 +08:00

distributed

Update pre-commit hooks (#12475)

2025-01-27 17:23:08 -07:00

engine

[Attention] MLA decode optimizations (#12528)

2025-01-30 23:49:37 -08:00

entrypoints

[Misc] fix typo: add missing space in lora adapter error message (#12564)

2025-01-30 15:39:22 +00:00

executor

[core] add wake_up doc and some sanity check (#12361)

2025-01-24 02:00:50 +08:00

inputs

[Misc] Rename MultiModalInputsV2 -> MultiModalInputs (#12244)

2025-01-21 07:31:19 +00:00

logging_utils

Rename vllm.logging to vllm.logging_utils (#10134)

2024-11-08 20:53:24 +00:00

lora

Update pre-commit hooks (#12475)

2025-01-27 17:23:08 -07:00

model_executor

[Feature] Fix guided decoding blocking bitmask memcpy (#12563)

2025-01-31 15:37:30 -08:00

multimodal

[Bugfix] Fix broken internvl2 inference with v1 (#12360)

2025-01-23 17:20:33 +00:00

platforms

[Attention] MLA decode optimizations (#12528)

2025-01-30 23:49:37 -08:00

plugins

[core][bugfix] configure env var during import vllm (#12209)

2025-01-20 19:35:59 +08:00

profiler

[Bugfix] Fix incorrect types in LayerwiseProfileResults (#12196)

2025-01-20 14:59:20 +08:00

prompt_adapter

Set weights_only=True when using torch.load() (#12366)

2025-01-24 02:17:30 +00:00

spec_decode

[CI] fix pre-commit error (#12494)

2025-01-28 06:11:05 +00:00

transformers_utils

[Bugfix] Gracefully handle huggingface hub http error (#12571)

2025-01-31 08:19:35 +00:00

triton_utils

[Docker] bump up neuron sdk v2.21 (#11593)

2024-12-30 13:46:14 +08:00

usage

[misc] add cuda runtime version to usage data (#12190)

2025-01-21 00:31:01 +00:00

v1

[v1][Bugfix] Add extra_keys to block_hash for prefix caching (#12603)

2025-01-31 13:13:04 -08:00

vllm_flash_attn

[ci][build] fix vllm-flash-attn (#8699)

2024-09-21 23:24:58 -07:00

worker

[Attention] MLA decode optimizations (#12528)

2025-01-30 23:49:37 -08:00

__init__.py

[core][bugfix] configure env var during import vllm (#12209)

2025-01-20 19:35:59 +08:00

_custom_ops.py

[Kernel][Quantization] Integrate block-quantized CUTLASS kernels for DeepSeekV3 (#12587)

2025-01-31 15:29:11 -08:00

_ipex_ops.py

[Misc][XPU] Upgrade to Pytorch 2.5 for xpu backend (#9823)

2024-11-06 17:29:03 -08:00

beam_search.py

[Frontend] re-enable multi-modality input in the new beam search implementation (#9427)

2024-10-29 11:49:47 +00:00

config.py

[Attention] MLA decode optimizations (#12528)

2025-01-30 23:49:37 -08:00

connections.py

Misc: allow to use proxy in HTTPConnection (#12042)

2025-01-15 13:16:40 +00:00

envs.py

[Attention] MLA decode optimizations (#12528)

2025-01-30 23:49:37 -08:00

forward_context.py

[torch.compile] Hide KV cache behind torch.compile boundary (#11677)

2025-01-10 13:14:42 +08:00

logger.py

[Misc] Move print_*_once from utils to logger (#11298)

2025-01-09 12:48:12 +08:00

logits_process.py

[Frontend] Bad words sampling parameter (#9717)

2024-10-26 16:29:38 +00:00

outputs.py

[V1][Frontend] Coalesce bunched RequestOutputs (#12298)

2025-01-23 17:17:41 -08:00

pooling_params.py

[Doc][4/N] Reorganize API Reference (#11843)

2025-01-08 21:34:44 +08:00

py.typed

Add py.typed so consumers of vLLM can get type checking (#1509)

2023-10-30 14:50:47 -07:00

sampling_params.py

[Bugfix] Fix OpenAI parallel sampling when using xgrammar (#11637)

2024-12-31 03:43:54 +00:00

scalar_type.py

Update pre-commit hooks (#12475)

2025-01-27 17:23:08 -07:00

scripts.py

[Frontend] Support reasoning content for deepseek r1 (#12473)

2025-01-29 11:38:08 +08:00

sequence.py

[Bugfix] Multi-sequence broken (#11898)

2025-01-21 11:51:35 -08:00

tracing.py

[Misc] Remove experimental dep from tracing.py (#12007)

2025-01-21 11:51:55 -08:00

utils.py

Update pre-commit hooks (#12475)

2025-01-27 17:23:08 -07:00

version.py

[CI/Build] use setuptools-scm to set __version__ (#4738)

2024-09-23 09:44:26 -07:00