Kunshang Ji
64b77dfd7e
[Core]fix type annotation for swap_blocks ( #4726 )
2024-05-10 21:52:48 +09:00
youkaichao
208b71bcc1
[Core][Distributed] refactor pynccl ( #4591 )
...
[Core][Distributed] refactor pynccl to hold multiple communicators (#4591 )
2024-05-09 19:48:43 -07:00
Cody Yu
c833101740
[Kernel] Refactor FP8 kv-cache with NVIDIA float8_e4m3 support ( #4535 )
2024-05-09 18:04:17 -06:00
Philipp Moritz
379da6dcb5
[Kernel] [FP8] Improve FP8 linear layer performance ( #4691 )
...
This PR improves the FP8 performance of linear layers, which had been lacking before (#4118 (comment) and #4118 (comment)).
We noticed that CUBLASLt can find a better algorithm if the first dimension of the matrix is greater than 16. So this PR enlarges matrices appropriately during quantization. This improves FP8 performance and removes the performance regression vs. FP16, in many cases exceeding FP16 performance.
Here are benchmarks on llama3 70b (ITL numbers for 1000 input and 50 output tokens at fixed qps and at TP 4), all FP8 measurements are for dynamic quantization:
qps = 1: 24 ms (FP8, this PR), 32 ms (FP8, previous main), 26 ms (FP16)
qps = 2: 26 ms (FP8, this PR), 34ms (FP8, previous main), 28 ms (FP16)
qps = 4: 33 ms (FP8, this PR), 44 ms (FP8, previous main), 36 ms (FP16)
qps = 6: 46 ms (FP8, this PR), 56 ms (FP8, previous main), 54 ms (FP16)
qps = 8: 85 ms (FP8, this PR), 85 ms (FP8, previous main), 138 ms (FP16)
2024-05-09 16:38:07 -07:00
Hao Zhang
ebce310b74
[Model] Snowflake arctic model implementation ( #4652 )
...
Co-authored-by: Dash Desai <1723932+iamontheinet@users.noreply.github.com >
Co-authored-by: Aurick Qiao <qiao@aurick.net >
Co-authored-by: Aurick Qiao <aurick.qiao@snowflake.com >
Co-authored-by: Aurick Qiao <aurickq@users.noreply.github.com >
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com >
2024-05-09 22:37:14 +00:00
Michael Goin
be0c5180ac
[Bugfix] Add logs for all model dtype casting ( #4717 )
2024-05-09 18:36:25 +00:00
Woosuk Kwon
0ee535b294
[Misc] Set block size at initialization & Fix test_model_runner ( #4705 )
2024-05-09 09:04:59 -07:00
Cyrus Leung
f12b20decc
[Frontend] Move async logic outside of constructor ( #4674 )
2024-05-08 22:48:33 -07:00
Mahmoud Ashraf
16bc0a098f
[Frontend] add tok/s speed metric to llm class when using tqdm ( #4400 )
...
Co-authored-by: Michael Goin <michael@neuralmagic.com >
2024-05-08 22:02:31 -07:00
Cade Daniel
8b9241be3a
[Speculative decoding] [Bugfix] Fix overallocation in ngram + spec logprobs ( #4672 )
2024-05-08 23:24:46 +00:00
Cody Yu
f942efb5a3
[Dynamic Spec Decoding] Auto-disable by the running queue size ( #4592 )
...
Co-authored-by: Cade Daniel <edacih@gmail.com >
2024-05-08 21:44:00 +00:00
Woosuk Kwon
89579a201f
[Misc] Use vllm-flash-attn instead of flash-attn ( #4686 )
2024-05-08 13:15:34 -07:00
youkaichao
20cfcdec99
[Core][Optimization] change python dict to pytorch tensor for blocks to swap ( #4659 )
2024-05-08 12:07:05 -07:00
Antoni Baum
ad932a221d
[Core] Faster startup for LoRA enabled models ( #4634 )
2024-05-08 10:33:18 -07:00
Woosuk Kwon
5510cf0e8a
[Misc] Add get_name method to attention backends ( #4685 )
2024-05-08 09:59:31 -07:00
DefTruth
0f9a6e3d22
[Bugfix][Kernel] allow non-power-of-2 for prefix prefill with alibi ( #4573 )
2024-05-08 09:19:58 -07:00
SangBin Cho
f6a593093a
[CI] Make mistral tests pass ( #4596 )
2024-05-08 08:44:35 -07:00
SangBin Cho
d7740ea4dc
[Core] Optimize sampler get_logprobs ( #4594 )
2024-05-08 08:42:28 -07:00
youkaichao
cc466a3290
[Core][Distributed] support cpu&device in broadcast tensor dict ( #4660 )
...
[Core][Distributed] support both cpu and device tensor in broadcast tensor dict (#4660 )
2024-05-07 19:34:47 -07:00
leiwen83
8344f7742b
[Bug fix][Core] fixup ngram not setup correctly ( #4551 )
...
Co-authored-by: Lei Wen <wenlei03@qiyi.com >
Co-authored-by: Cade Daniel <edacih@gmail.com >
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com >
2024-05-07 11:40:18 -07:00
youkaichao
469f85c782
[Core][Optimization] change copy-on-write from dict[int, list] to list ( #4648 )
2024-05-07 11:06:32 -07:00
Austin Veselka
10760da800
[Bugfix] Fixed error in slice_lora_b for MergedQKVParallelLinearWithLora ( #4609 )
2024-05-07 10:59:07 -07:00
youkaichao
63575bc2e1
[Core][Optimization] change python dict to pytorch tensor ( #4607 )
2024-05-06 21:30:27 -07:00
Cyrus Leung
323f27b904
[Bugfix] Fix asyncio.Task not being subscriptable ( #4623 )
2024-05-06 09:31:05 -07:00
zhaoyang-star
0650e5935b
Disable cuda version check in vllm-openai image ( #4530 )
2024-05-05 16:58:55 -07:00
Simon Mo
8d8357c8ed
bump version to v0.4.2 ( #4600 )
2024-05-04 17:09:49 -07:00
DearPlanet
4302987069
[Bugfix] Fix inappropriate content of model_name tag in Prometheus metrics ( #3937 )
2024-05-04 15:39:34 -07:00
Michael Goin
2a052011ca
[Kernel] Support MoE Fp8 Checkpoints for Mixtral (Static Weights with Dynamic/Static Activations) ( #4527 )
...
Follow on to #4332 to enable FP8 checkpoint loading for Mixtral and supersedes #4436 .
This PR enables the following checkpoint loading features for Mixtral:
Supports loading fp8 checkpoints for Mixtral, such as this "nm-testing/Mixtral-8x7B-Instruct-v0.1-FP8" test model
Supports static or dynamic activation quantization with static weight quantization (all per tensor)
Supports different scales for each expert weight
Supports Fp8 in QKV layer
Notes:
The Expert Gate/Router always runs at half / full precision for now.
If there are different weight scales between QKV layer (for separate QKV weights), they are re-quantized using layer.weight_scale.max() so we can have a single gemm for performance.
2024-05-04 11:45:16 -07:00
SangBin Cho
36fb68f947
[Doc] Chunked Prefill Documentation ( #4580 )
2024-05-04 00:18:00 -07:00
Cody Yu
bc8ad68455
[Misc][Refactor] Introduce ExecuteModelData ( #4540 )
2024-05-03 17:47:07 -07:00
youkaichao
344bf7cd2d
[Misc] add installation time env vars ( #4574 )
2024-05-03 15:55:56 -07:00
Cade Daniel
ab50275111
[Speculative decoding] Support target-model logprobs ( #4378 )
2024-05-03 15:52:01 -07:00
Lily Liu
43c413ec57
[Kernel] Use flashinfer for decoding ( #4353 )
...
Co-authored-by: LiuXiaoxuanPKU <llilyliupku@gmail.com >
2024-05-03 15:51:27 -07:00
Sebastian Schoennenbeck
f8e7adda21
Fix/async chat serving ( #2727 )
2024-05-03 11:04:14 -07:00
Michael Goin
7e65477e5e
[Bugfix] Allow "None" or "" to be passed to CLI for string args that default to None ( #4586 )
2024-05-03 10:32:21 -07:00
SangBin Cho
3521ba4f25
[Core][Model runner refactoring 1/N] Refactor attn metadata term ( #4518 )
2024-05-03 10:20:12 -07:00
youkaichao
2d7bce9cd5
[Doc] add env vars to the doc ( #4572 )
2024-05-03 05:13:49 +00:00
DefTruth
ce3f1eedf8
[Misc] remove chunk detected debug logs ( #4571 )
2024-05-03 04:48:08 +00:00
Yang, Bo
808632d3b4
[BugFix] Prevent the task of _force_log from being garbage collected ( #4567 )
2024-05-03 01:35:18 +00:00
youkaichao
344a5d0c33
[Core][Distributed] enable allreduce for multiple tp groups ( #4566 )
2024-05-02 17:32:33 -07:00
SangBin Cho
0f8a91401c
[Core] Ignore infeasible swap requests. ( #4557 )
2024-05-02 14:31:20 -07:00
Michał Moskal
32881f3f31
[kernel] fix sliding window in prefix prefill Triton kernel ( #4405 )
...
Co-authored-by: SangBin Cho <rkooo567@gmail.com >
2024-05-02 11:23:37 -07:00
youkaichao
5b8a7c1cb0
[Misc] centralize all usage of environment variables ( #4548 )
2024-05-02 11:13:25 -07:00
alexm-nm
7038e8b803
[Kernel] Support running GPTQ 8-bit models in Marlin ( #4533 )
2024-05-02 12:56:22 -04:00
youkaichao
2a85f93007
[Core][Distributed] enable multiple tp group ( #4512 )
...
Co-authored-by: Zhuohan Li <zhuohan123@gmail.com >
2024-05-02 04:28:21 +00:00
SangBin Cho
cf8cac8c70
[mypy][6/N] Fix all the core subdirectory typing ( #4450 )
...
Co-authored-by: Cade Daniel <edacih@gmail.com >
2024-05-02 03:01:00 +00:00
SangBin Cho
0d62fe58db
[Bug fix][Core] assert num_new_tokens == 1 fails when SamplingParams.n is not 1 and max_tokens is large & Add tests for preemption ( #4451 )
2024-05-01 19:24:13 -07:00
Danny Guinther
b8afa8b95a
[MISC] Rework logger to enable pythonic custom logging configuration to be provided ( #4273 )
2024-05-01 17:34:40 -07:00
Woosuk Kwon
826b82a260
[Misc] Fix expert_ids shape in MoE ( #4517 )
2024-05-01 23:47:59 +00:00
Philipp Moritz
c9d852d601
[Misc] Remove Mixtral device="cuda" declarations ( #4543 )
...
Remove the device="cuda" declarations in mixtral as promised in #4343
2024-05-01 16:30:52 -07:00
