Woosuk Kwon
41ca62cf03
[Misc] Add CustomOp interface for device portability ( #5255 )
2024-06-05 09:18:19 -07:00
Woosuk Kwon
3a434b07ed
[Kernel] Enhance MoE benchmarking & tuning script ( #4921 )
2024-06-03 20:06:59 -07:00
Toshiki Kataoka
06b2550cbb
[Bugfix] Support prompt_logprobs==0 ( #5217 )
2024-06-03 17:59:30 -07:00
Breno Faria
f775a07e30
[FRONTEND] OpenAI tools support named functions ( #5032 )
2024-06-03 18:25:29 -05:00
Tyler Michael Smith
cbb2f59cc8
[Kernel] Pass a device pointer into the quantize kernel for the scales ( #5159 )
2024-06-03 09:52:30 -07:00
Cyrus Leung
7a64d24aad
[Core] Support image processor ( #4197 )
2024-06-02 22:56:41 -07:00
Divakar Verma
a66cf40b20
[Kernel][ROCm][AMD] enable fused topk_softmax kernel for moe layer ( #4927 )
...
This PR enables the fused topk_softmax kernel used in moe layer for HIP
2024-06-02 14:13:26 -07:00
chenqianfzh
b9c0605a8e
[Feature][Kernel] Support bitsandbytes quantization and QLoRA ( #4776 )
2024-06-01 14:51:10 -06:00
Ye Cao
c354072828
[Minor] Fix the path typo in loader.py: save_sharded_states.py -> save_sharded_state.py ( #5151 )
...
Signed-off-by: Ye Cao <caoye.cao@alibaba-inc.com >
2024-06-01 17:11:22 +00:00
Tyler Michael Smith
260d119e86
[Kernel] Refactor CUTLASS kernels to always take scales that reside on the GPU ( #5137 )
2024-06-01 06:45:32 +00:00
Cody Yu
e9899fb7a4
[Model] Enable FP8 QKV in MoE and refine kernel tuning script ( #5039 )
2024-05-31 14:29:19 -07:00
Robert Shaw
b35be5403f
[Bugfix] Avoid Warnings in SparseML Activation Quantization ( #5120 )
2024-05-30 17:04:37 -07:00
Alexander Matveev
5bf185a1c4
[Bugfix] gptq_marlin: Ensure g_idx_sort_indices is not a Parameter ( #5108 )
2024-05-30 00:30:18 +00:00
Divakar Verma
dd8de11f0a
[Kernel][ROCm][AMD] Add fused_moe Triton configs for MI300X ( #4951 )
...
This PR adds Triton kernel configs for the MoE kernel for MI300X
2024-05-28 16:03:23 +00:00
Isotr0py
890aa93d27
[Model] Add support for falcon-11B ( #5069 )
2024-05-27 16:41:43 -07:00
sasha0552
fbdb7b3ee2
[Core] Allow AQLM on Pascal ( #5058 )
2024-05-27 15:26:14 -07:00
1102bef219
[Bugfix / Core] Prefix Caching Guards (merged with main) ( #4846 )
...
Co-authored-by: rsnm2 <rshaw@neuralmagic.com >
Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com >
2024-05-27 15:18:17 -07:00
8e192ff967
[Kernel][Backend][Model] Blocksparse flash attention kernel and Phi-3-Small model ( #4799 )
...
Co-authored-by: beagleski <yunanzhang@microsoft.com >
Co-authored-by: bapatra <bapatra@microsoft.com >
Co-authored-by: Barun Patra <codedecde@users.noreply.github.com >
Co-authored-by: Michael Goin <michael@neuralmagic.com >
2024-05-24 22:00:52 -07:00
Robert Shaw and Cody Yu
919770957f
[Bugfix] Fix Mistral v0.3 Weight Loading ( #5005 )
...
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com >
2024-05-24 12:28:27 +00:00
Elisei Smirnov and Elisei Smirnov
e3470f8753
[Core]: Option To Use Prompt Token Ids Inside Logits Processor ( #4985 )
...
Co-authored-by: Elisei Smirnov <el.smirnov@innopolis.university >
2024-05-23 22:04:24 +00:00
a1242324c9
[Kernel] Initial Activation Quantization Support ( #4525 )
...
Co-authored-by: Varun Sundar Rabindranath <varunsundar08@gmail.com >
Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com >
2024-05-23 21:29:18 +00:00
Alexander Matveev
6066253296
Marlin 24 prefill performance improvement (about 25% better on average) ( #4983 )
2024-05-23 02:39:27 -04:00
Philipp Moritz
a36de682d4
[Minor] Fix small typo in llama.py: QKVParallelLinear -> QuantizationConfig ( #4991 )
2024-05-22 22:26:56 +00:00
raywanb
97b030005c
[Model] LoRA gptbigcode implementation ( #3949 )
2024-05-22 13:58:59 -07:00
Cody Yu
a3a73ab069
[Misc] Load FP8 kv-cache scaling factors from checkpoints ( #4893 )
...
The 2nd PR for #4532 .
This PR supports loading FP8 kv-cache scaling factors from a FP8 checkpoint (with .kv_scale parameter).
2024-05-22 13:28:20 -07:00
Isotr0py
f12c3b5b3d
[Model] Add Phi-2 LoRA support ( #4886 )
2024-05-21 14:24:17 +09:00
HUANG Fei
d130b573a0
[Model] add rope_scaling support for qwen2 ( #4930 )
2024-05-21 05:22:22 +00:00
Aurick Qiao
1937e29848
[Core] Sharded State Loader download from HF ( #4889 )
2024-05-20 11:46:12 -07:00
Mor Zusman and Mor Zusman
f0eecee610
[Bugfix] Fix dummy weight for fp8 ( #4916 )
...
Allow dummy load format for fp8,
torch.uniform_ doesn't support FP8 at the moment
Co-authored-by: Mor Zusman <morz@ai21.com >
2024-05-20 18:44:25 +00:00
Cyrus Leung
6287537a0c
[Model] LLaVA model refactor ( #4910 )
2024-05-20 08:11:25 +00:00
Alexander Matveev
27ce85476e
[Kernel] Add marlin_24 unit tests ( #4901 )
2024-05-19 11:37:34 -04:00
Cyrus Leung
f68470e803
[Bugfix][Model] Add base class for vision-language models ( #4809 )
2024-05-19 00:13:33 -07:00
SangBin Cho
2e9a2227ec
[Lora] Support long context lora ( #4787 )
...
Currently we need to call rotary embedding kernel for each LoRA, which makes it hard to serve multiple long context length LoRA. Add batched rotary embedding kernel and pipe it through.
It replaces the rotary embedding layer to the one that is aware of multiple cos-sin-cache per scaling factors.
Follow up of https://github.com/vllm-project/vllm/pull/3095/files
2024-05-18 16:05:23 +09:00
eigenLiu
48d5985a08
Sync huggingface modifications of qwen Moe model ( #4774 )
2024-05-17 09:43:19 -07:00
Jinzhen Lin
33e0823de5
[Bugfix] fix rope error when load models with different dtypes ( #4835 )
2024-05-17 18:43:34 +09:00
Alexander Matveev and Robert Shaw
6979ade384
Add GPTQ Marlin 2:4 sparse structured support ( #4790 )
...
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com >
2024-05-16 12:56:15 -04:00
Jinzhen Lin
99caa49106
[Kernel] add bfloat16 support for gptq marlin kernel ( #4788 )
2024-05-16 09:55:29 -04:00
alexm-nm
5c342570d7
Add marlin unit tests and marlin benchmark script ( #4815 )
2024-05-16 09:36:49 -04:00
Aurick Qiao and Woosuk Kwon
30e754390c
[Core] Implement sharded state loader ( #4690 )
...
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu >
2024-05-15 22:11:54 -07:00
SangBin Cho
65bf2ac165
[Core][2/N] Model runner refactoring part 2. Combine prepare prefill / decode to a single API ( #4681 )
...
This PR combines prepare_prompt and prepare_decode into a single API. This PR also coelsce the attn metadata for prefill/decode to a single class and allow to slice them when running attn backend.
It also refactors subquery_start_loc which was not refactored in the previous PR
2024-05-15 14:00:10 +09:00
Philipp Moritz
33d3914b1e
[Bugfix] Fix dynamic FP8 quantization for Mixtral ( #4793 )
2024-05-13 19:00:27 -04:00
Sanger Steel
8bc68e198c
[Frontend] [Core] perf: Automatically detect vLLM-tensorized model, update tensorizer to version 2.9.0 ( #4208 )
2024-05-13 14:57:07 -07:00
Woosuk Kwon
0fca3cdcf2
[Misc] Enhance attention selector ( #4751 )
2024-05-13 10:47:25 -07:00
Swapnil Parekh
a7be4d0072
[CORE] Improvement in ranks code ( #4718 )
2024-05-12 17:47:47 -07:00
Yikang Shen
6eaccb7353
[Model] Add support for IBM Granite Code models ( #4636 )
2024-05-11 21:27:24 -07:00
Chang Su
e254497b66
[Model][Misc] Add e5-mistral-7b-instruct and Embedding API ( #3734 )
2024-05-11 11:30:37 -07:00
SangBin Cho
6a0f617210
[Core] Fix circular reference which leaked llm instance in local dev env ( #4737 )
...
Storing exception frame is extremely prone to circular refernece because it contains the reference to objects.
When tensorizer is not installed, it leaks llm instance because error frame has references to various modules which cause circular reference problem.
I also found spec decoding has a circular reference issue, and I solved it using weakref.proxy.
2024-05-10 23:54:32 +09:00
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
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
Cody Yu and Cade Daniel
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