[Kernel] Enable 8-bit weights in Fused Marlin MoE #8032

ElizaWszola · 2024-08-30T13:14:26Z

This PR covers the remaining work of PR #7079 . It enables 8-bit weights in the fused Marlin kernel without modifying the codepaths that run the kernel and also adds 8-bit tests to test_moe.py.

Tested e2e by running offline_inference.py with

llm = LLM(model="nm-testing/Mixtral-8x7B-Instruct-v0.1-W8A16-quantized")

and

llm = LLM(model="TheBloke/Mixtral-8x7B-v0.1-GPTQ",
          revision="gptq-8bit-128g-actorder_True")

Sonnet benchmark results (with act order, 8-bit):

TTFT	TPOT

github-actions · 2024-08-30T13:14:38Z

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

Comment /ready on the PR
Add ready label to the PR
Enable auto-merge.

🚀

dsikka

Just left one quick comment. I'm going to pull this PR in and try it with a compressed-tensors W8A16 model.

dsikka · 2024-08-30T16:01:04Z

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

+        from vllm.model_executor.layers.fused_moe.fused_moe_marlin import (
+            fused_moe_marlin)
+
+        return fused_moe_marlin(


could we keep the name the same?

dsikka · 2024-08-30T16:05:58Z

/ready

dsikka · 2024-08-30T17:03:01Z

Just left one quick comment. I'm going to pull this PR in and try it with a compressed-tensors W8A16 model.

Confirmed this works with compressed-tensors w8a16

dsikka · 2024-08-30T20:17:52Z

vllm/model_executor/layers/fused_moe/fused_moe_marlin.py

+    else:
+        topk_weights, topk_ids = custom_routing_function(
+            hidden_states, gating_output, topk, renormalize)
+


Instead of doing this here, could we use FusedMoe.select_experts?
Similar to what fp8 does:

vllm/vllm/model_executor/layers/quantization/fp8.py

Line 486 in 1248e85

topk_weights, topk_ids = FusedMoE.select_experts(

In that case, the topk_weights and topk_ids would both be passed to fused_moe_marlin

I think this would also be needed for deepseek-v2 which requires grouped topk

dsikka · 2024-08-30T20:26:44Z

Still need to test with a deepseek-v2 model

csrc/moe/marlin_moe_ops.cu

dsikka · 2024-09-04T21:04:00Z

@ElizaWszola seems like the kernel test failures start after tests/kernels/test_moe.py - could you take a look?

dsikka

This LGTM pending the deepseek int4 model/group topk works fine.
I'll quantize a deepseek int8 model over the weekend and we can verify with that as well.

Thanks!

dsikka · 2024-09-13T16:17:56Z

tests/weight_loading/run_model_weight_loading_test.sh

What's going on this with this file?

I made some accidental changes in it when merging, and when I reverted them, it just showed no lines in diff

I don't know why it's still there...

mgoin

@dsikka confirmed that Deepseek with group quantization works, so all good! Nice work

Co-authored-by: Dipika <dipikasikka1@gmail.com>

* [Kernel] Enable 8-bit weights in Fused Marlin MoE (vllm-project#8032) Co-authored-by: Dipika <dipikasikka1@gmail.com> * [Frontend] Expose revision arg in OpenAI server (vllm-project#8501) * [BugFix] Fix clean shutdown issues (vllm-project#8492) * [Bugfix][Kernel] Fix build for sm_60 in GGUF kernel (vllm-project#8506) * [Kernel] AQ AZP 3/4: Asymmetric quantization kernels (vllm-project#7270) * [doc] update doc on testing and debugging (vllm-project#8514) * [Bugfix] Bind api server port before starting engine (vllm-project#8491) * [perf bench] set timeout to debug hanging (vllm-project#8516) * [misc] small qol fixes for release process (vllm-project#8517) * [Bugfix] Fix 3.12 builds on main (vllm-project#8510) Signed-off-by: Joe Runde <Joseph.Runde@ibm.com> * [refactor] remove triton based sampler (vllm-project#8524) * [Frontend] Improve Nullable kv Arg Parsing (vllm-project#8525) Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com> * [Misc][Bugfix] Disable guided decoding for mistral tokenizer (vllm-project#8521) * [torch.compile] register allreduce operations as custom ops (vllm-project#8526) * [Misc] Limit to ray[adag] 2.35 to avoid backward incompatible change (vllm-project#8509) Signed-off-by: Rui Qiao <ruisearch42@gmail.com> * [Benchmark] Support sample from HF datasets and image input for benchmark_serving (vllm-project#8495) * [Encoder decoder] Add cuda graph support during decoding for encoder-decoder models (vllm-project#7631) * [Feature][kernel] tensor parallelism with bitsandbytes quantization (vllm-project#8434) * [Model] Add mistral function calling format to all models loaded with "mistral" format (vllm-project#8515) Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com> * [Misc] Don't dump contents of kvcache tensors on errors (vllm-project#8527) * [Bugfix] Fix TP > 1 for new granite (vllm-project#8544) Signed-off-by: Joe Runde <Joseph.Runde@ibm.com> * [doc] improve installation doc (vllm-project#8550) Co-authored-by: Andy Dai <76841985+Imss27@users.noreply.github.com> * [CI/Build] Excluding kernels/test_gguf.py from ROCm (vllm-project#8520) * [Kernel] Change interface to Mamba causal_conv1d_update for continuous batching (vllm-project#8012) * [CI/Build] fix Dockerfile.cpu on podman (vllm-project#8540) * [Misc] Add argument to disable FastAPI docs (vllm-project#8554) * [CI/Build] Avoid CUDA initialization (vllm-project#8534) * [CI/Build] Update Ruff version (vllm-project#8469) Signed-off-by: Aaron Pham <contact@aarnphm.xyz> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com> * [Core][Bugfix][Perf] Introduce `MQLLMEngine` to avoid `asyncio` OH (vllm-project#8157) Co-authored-by: Nick Hill <nickhill@us.ibm.com> Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com> Co-authored-by: Simon Mo <simon.mo@hey.com> * [Core] *Prompt* logprobs support in Multi-step (vllm-project#8199) * [Core] zmq: bind only to 127.0.0.1 for local-only usage (vllm-project#8543) Signed-off-by: Russell Bryant <rbryant@redhat.com> * [Model] Support Solar Model (vllm-project#8386) Co-authored-by: Michael Goin <michael@neuralmagic.com> * [AMD][ROCm]Quantization methods on ROCm; Fix _scaled_mm call (vllm-project#8380) Co-authored-by: Alexei-V-Ivanov-AMD <156011006+Alexei-V-Ivanov-AMD@users.noreply.github.com> Co-authored-by: Michael Goin <michael@neuralmagic.com> * [Kernel] Change interface to Mamba selective_state_update for continuous batching (vllm-project#8039) * [BugFix] Nonzero exit code if MQLLMEngine startup fails (vllm-project#8572) * [Bugfix] add `dead_error` property to engine client (vllm-project#8574) Signed-off-by: Joe Runde <Joseph.Runde@ibm.com> * [Kernel] Remove marlin moe templating on thread_m_blocks (vllm-project#8573) Co-authored-by: lwilkinson@neuralmagic.com * [Bugfix] [Encoder-Decoder] Bugfix for encoder specific metadata construction during decode of encoder-decoder models. (vllm-project#8545) * Revert "[Misc][Bugfix] Disable guided decoding for mistral tokenizer" (vllm-project#8593) * [Bugfix] fixing sonnet benchmark bug in benchmark_serving.py (vllm-project#8616) * [MISC] remove engine_use_ray in benchmark_throughput.py (vllm-project#8615) * [Frontend] Use MQLLMEngine for embeddings models too (vllm-project#8584) * [Kernel][Amd] Add fp8 kv cache support for rocm custom paged attention (vllm-project#8577) * [Core] simplify logits resort in _apply_top_k_top_p (vllm-project#8619) * [Doc] Add documentation for GGUF quantization (vllm-project#8618) * Create SECURITY.md (vllm-project#8642) * [CI/Build] Re-enabling Entrypoints tests on ROCm, excluding ones that fail (vllm-project#8551) * [Misc] guard against change in cuda library name (vllm-project#8609) * [Bugfix] Fix Phi3.5 mini and MoE LoRA inference (vllm-project#8571) * [bugfix] [AMD] add multi-step advance_step to ROCmFlashAttentionMetadata (vllm-project#8474) * [Core] Support Lora lineage and base model metadata management (vllm-project#6315) * [Model] Add OLMoE (vllm-project#7922) * [CI/Build] Removing entrypoints/openai/test_embedding.py test from ROCm build (vllm-project#8670) * [Bugfix] Validate SamplingParam n is an int (vllm-project#8548) * [Misc] Show AMD GPU topology in `collect_env.py` (vllm-project#8649) * [Bugfix] Config got an unexpected keyword argument 'engine' (vllm-project#8556) * [Bugfix][Core] Fix tekken edge case for mistral tokenizer (vllm-project#8640) * [Doc] neuron documentation update (vllm-project#8671) Signed-off-by: omrishiv <327609+omrishiv@users.noreply.github.com> * [Hardware][AWS] update neuron to 2.20 (vllm-project#8676) Signed-off-by: omrishiv <327609+omrishiv@users.noreply.github.com> * [Bugfix] Fix incorrect llava next feature size calculation (vllm-project#8496) * [Core] Rename `PromptInputs` and `inputs`(vllm-project#8673) * [MISC] add support custom_op check (vllm-project#8557) Co-authored-by: youkaichao <youkaichao@126.com> * [Core] Factor out common code in `SequenceData` and `Sequence` (vllm-project#8675) * [beam search] add output for manually checking the correctness (vllm-project#8684) * [Kernel] Build flash-attn from source (vllm-project#8245) * [VLM] Use `SequenceData.from_token_counts` to create dummy data (vllm-project#8687) * [Doc] Fix typo in AMD installation guide (vllm-project#8689) * [Kernel][Triton][AMD] Remove tl.atomic_add from awq_gemm_kernel, 2-5x speedup MI300, minor improvement for MI250 (vllm-project#8646) * [dbrx] refactor dbrx experts to extend FusedMoe class (vllm-project#8518) * [Kernel][Bugfix] Delete some more useless code in marlin_moe_ops.cu (vllm-project#8643) * [Bugfix] Refactor composite weight loading logic (vllm-project#8656) * [ci][build] fix vllm-flash-attn (vllm-project#8699) * [Model] Refactor BLIP/BLIP-2 to support composite model loading (vllm-project#8407) * [Misc] Use NamedTuple in Multi-image example (vllm-project#8705) Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com> * [MISC] rename CudaMemoryProfiler to DeviceMemoryProfiler (vllm-project#8703) * [Model][VLM] Add LLaVA-Onevision model support (vllm-project#8486) Co-authored-by: litianjian <litianjian@bytedance.com> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com> Co-authored-by: Roger Wang <ywang@roblox.com> Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk> * [SpecDec][Misc] Cleanup, remove bonus token logic. (vllm-project#8701) * [build] enable existing pytorch (for GH200, aarch64, nightly) (vllm-project#8713) * [misc] upgrade mistral-common (vllm-project#8715) * [Bugfix] Avoid some bogus messages RE CUTLASS's revision when building (vllm-project#8702) * [Bugfix] Fix CPU CMake build (vllm-project#8723) Co-authored-by: Yuan <yuan.zhou@intel.com> * [Bugfix] fix docker build for xpu (vllm-project#8652) * [Core][Frontend] Support Passing Multimodal Processor Kwargs (vllm-project#8657) Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com> * [Hardware][CPU] Refactor CPU model runner (vllm-project#8729) * [Bugfix][CPU] fix missing input intermediate_tensors in the cpu_model_runner (vllm-project#8733) * [Model] Support pp for qwen2-vl (vllm-project#8696) * [VLM] Fix paligemma, fuyu and persimmon with transformers 4.45 : use config.text_config.vocab_size (vllm-project#8707) * [CI/Build] use setuptools-scm to set __version__ (vllm-project#4738) Co-authored-by: youkaichao <youkaichao@126.com> * [Kernel] (2/N) Machete - Integrate into CompressedTensorsWNA16 and GPTQMarlin (vllm-project#7701) Co-authored-by: mgoin <michael@neuralmagic.com> Co-authored-by: Divakar Verma <137818590+divakar-amd@users.noreply.github.com> Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com> * [Kernel][LoRA] Add assertion for punica sgmv kernels (vllm-project#7585) * [Core] Allow IPv6 in VLLM_HOST_IP with zmq (vllm-project#8575) Signed-off-by: Russell Bryant <rbryant@redhat.com> * Fix typical acceptance sampler with correct recovered token ids (vllm-project#8562) * Add output streaming support to multi-step + async while ensuring RequestOutput obj reuse (vllm-project#8335) * [Hardware][AMD] ROCm6.2 upgrade (vllm-project#8674) * Fix tests in test_scheduler.py that fail with BlockManager V2 (vllm-project#8728) * re-implement beam search on top of vllm core (vllm-project#8726) Co-authored-by: Brendan Wong <bjwpokemon@gmail.com> * Revert "[Core] Rename `PromptInputs` to `PromptType`, and `inputs` to `prompt`" (vllm-project#8750) * [MISC] Skip dumping inputs when unpicklable (vllm-project#8744) * [Core][Model] Support loading weights by ID within models (vllm-project#7931) * [Model] Expose Phi3v num_crops as a mm_processor_kwarg (vllm-project#8658) Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com> Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk> * [Bugfix] Fix potentially unsafe custom allreduce synchronization (vllm-project#8558) * [Kernel] Split Marlin MoE kernels into multiple files (vllm-project#8661) Co-authored-by: mgoin <michael@neuralmagic.com> * [Frontend] Batch inference for llm.chat() API (vllm-project#8648) Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk> Co-authored-by: Roger Wang <ywang@roblox.com> Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com> * [Bugfix] Fix torch dynamo fixes caused by `replace_parameters` (vllm-project#8748) * [CI/Build] fix setuptools-scm usage (vllm-project#8771) * [misc] soft drop beam search (vllm-project#8763) * [[Misc]Upgrade bitsandbytes to the latest version 0.44.0 (vllm-project#8768) * [Core][Bugfix] Support prompt_logprobs returned with speculative decoding (vllm-project#8047) Signed-off-by: Travis Johnson <tsjohnso@us.ibm.com> * [Core] Adding Priority Scheduling (vllm-project#5958) * [Bugfix] Use heartbeats instead of health checks (vllm-project#8583) * Fix test_schedule_swapped_simple in test_scheduler.py (vllm-project#8780) * [Bugfix][Kernel] Implement acquire/release polyfill for Pascal (vllm-project#8776) * Fix tests in test_chunked_prefill_scheduler which fail with BlockManager V2 (vllm-project#8752) * [BugFix] Propagate 'trust_remote_code' setting in internvl and minicpmv (vllm-project#8250) * [Hardware][CPU] Enable mrope and support Qwen2-VL on CPU backend (vllm-project#8770) * [Bugfix] load fc bias from config for eagle (vllm-project#8790) --------- Signed-off-by: Joe Runde <Joseph.Runde@ibm.com> Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com> Signed-off-by: Rui Qiao <ruisearch42@gmail.com> Signed-off-by: Aaron Pham <contact@aarnphm.xyz> Signed-off-by: Russell Bryant <rbryant@redhat.com> Signed-off-by: omrishiv <327609+omrishiv@users.noreply.github.com> Signed-off-by: Travis Johnson <tsjohnso@us.ibm.com> Co-authored-by: ElizaWszola <eliza@neuralmagic.com> Co-authored-by: Dipika <dipikasikka1@gmail.com> Co-authored-by: lewtun <lewis.c.tunstall@gmail.com> Co-authored-by: Nick Hill <nickhill@us.ibm.com> Co-authored-by: sasha0552 <admin@sasha0552.org> Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com> Co-authored-by: youkaichao <youkaichao@gmail.com> Co-authored-by: Kevin Lin <42618777+kevin314@users.noreply.github.com> Co-authored-by: Simon Mo <simon.mo@hey.com> Co-authored-by: Joe Runde <Joseph.Runde@ibm.com> Co-authored-by: Alex Brooks <alex.brooks@ibm.com> Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com> Co-authored-by: Rui Qiao <161574667+ruisearch42@users.noreply.github.com> Co-authored-by: Isotr0py <2037008807@qq.com> Co-authored-by: sroy745 <142070531+sroy745@users.noreply.github.com> Co-authored-by: chenqianfzh <51831990+chenqianfzh@users.noreply.github.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com> Co-authored-by: Andy Dai <76841985+Imss27@users.noreply.github.com> Co-authored-by: Alexey Kondratiev(AMD) <143633163+alexeykondrat@users.noreply.github.com> Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com> Co-authored-by: Daniele <36171005+dtrifiro@users.noreply.github.com> Co-authored-by: Jiaxin Shan <seedjeffwan@gmail.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk> Co-authored-by: Aaron Pham <contact@aarnphm.xyz> Co-authored-by: Alexander Matveev <59768536+alexm-neuralmagic@users.noreply.github.com> Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com> Co-authored-by: afeldman-nm <156691304+afeldman-nm@users.noreply.github.com> Co-authored-by: Russell Bryant <rbryant@redhat.com> Co-authored-by: Geun, Lim <shing100@Naver.com> Co-authored-by: Michael Goin <michael@neuralmagic.com> Co-authored-by: Gregory Shtrasberg <156009573+gshtras@users.noreply.github.com> Co-authored-by: Alexei-V-Ivanov-AMD <156011006+Alexei-V-Ivanov-AMD@users.noreply.github.com> Co-authored-by: Kuntai Du <kuntai@uchicago.edu> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Charlie Fu <charlifu@amd.com> Co-authored-by: 盏一 <w@hidva.com> Co-authored-by: bnellnm <49004751+bnellnm@users.noreply.github.com> Co-authored-by: Amit Garg <mitgarg17495@gmail.com> Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com> Co-authored-by: Niklas Muennighoff <n.muennighoff@gmail.com> Co-authored-by: saumya-saran <saumya.saran@c3.ai> Co-authored-by: Pastel！ <1627301104@qq.com> Co-authored-by: omrishiv <327609+omrishiv@users.noreply.github.com> Co-authored-by: zyddnys <zyddnys@outlook.com> Co-authored-by: youkaichao <youkaichao@126.com> Co-authored-by: rasmith <Randall.Smith@amd.com> Co-authored-by: Divakar Verma <137818590+divakar-amd@users.noreply.github.com> Co-authored-by: Huazhong Ji <hzji210@gmail.com> Co-authored-by: litianjian <45817262+litianjian@users.noreply.github.com> Co-authored-by: litianjian <litianjian@bytedance.com> Co-authored-by: Roger Wang <ywang@roblox.com> Co-authored-by: Lily Liu <lilyliupku@gmail.com> Co-authored-by: Yuan <yuan.zhou@intel.com> Co-authored-by: Yan Ma <yan.ma@intel.com> Co-authored-by: Li, Jiang <jiang1.li@intel.com> Co-authored-by: Yanyi Liu <wolfsonliu@163.com> Co-authored-by: Jani Monoses <jani.monoses@gmail.com> Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com> Co-authored-by: jiqing-feng <107918818+jiqing-feng@users.noreply.github.com> Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com> Co-authored-by: Brendan Wong <bjwpokemon@gmail.com> Co-authored-by: Cody Yu <hao.yu.cody@gmail.com> Co-authored-by: Peter Salas <peter@fixie.ai> Co-authored-by: Hanzhi Zhou <hanzhi713@gmail.com> Co-authored-by: Andy <37781802+aandyw@users.noreply.github.com> Co-authored-by: Travis Johnson <tsjohnso@us.ibm.com> Co-authored-by: Archit Patke <apatke@illinois.edu> Co-authored-by: zifeitong <zifeitong@gmail.com> Co-authored-by: sohamparikh <sohamparikh47@gmail.com>

Co-authored-by: Dipika <dipikasikka1@gmail.com>

Enable 8-bit weights in Fused Marlin MoE

0abac6f

ElizaWszola added 2 commits August 30, 2024 09:36

fix rocm

fdf69c2

bad paste

4da163b

dsikka reviewed Aug 30, 2024

View reviewed changes

github-actions bot added the ready ONLY add when PR is ready to merge/full CI is needed label Aug 30, 2024

dsikka added 3 commits August 30, 2024 17:29

add test case; fix imports for tests

21d2337

Merge branch 'main' into marlin-moe-8-bit

080ab23

fix to adapt custom_routin_function

638777a

dsikka reviewed Aug 30, 2024

View reviewed changes

ElizaWszola and others added 4 commits September 2, 2024 03:04

Use select_experts to compute top_k tensors in fused moe

bd4b84d

bring back fused_moe_marlin -> fused_marlin_moe

bef6b53

Merge branch 'main' into marlin-moe-8-bit

befc52b

remove large model

b45594c

mgoin reviewed Sep 4, 2024

View reviewed changes

csrc/moe/marlin_moe_ops.cu Outdated Show resolved Hide resolved

ElizaWszola added 2 commits September 4, 2024 13:10

Cleanup, comments

effd2cd

fix moe init

52c3353

dsikka and others added 9 commits September 4, 2024 21:14

move larger models to an options larger test

882fd9c

add optional flag

973d914

swap gpu

72bc899

Temp disable part of moe tests to see what's breaking

eea2bc3

Fixes to act_order, make unit tests more robust

9c29dc2

try to narrow down cuda error

6d04dcd

Try different subset of test params

83e7999

.

6a42eaf

.

3288842

ElizaWszola added 6 commits September 10, 2024 01:06

Merge branch 'main' into marlin-moe-8-bit

61ef4ba

fix and cleanup after merge

667d23e

cleanup

b16838e

validate cache for the kernel code

e53abb9

cleanup commented out code

2f82715

Merge branch 'main' into marlin-moe-8-bit

f97b524

dsikka reviewed Sep 13, 2024

View reviewed changes

ElizaWszola added 3 commits September 13, 2024 14:20

Merge branch 'main' into marlin-moe-8-bit

771f693

fix type conversion

aac7c20

Merge branch 'main' into marlin-moe-8-bit

9d7caad

mgoin approved these changes Sep 16, 2024

View reviewed changes

mgoin merged commit a091e2d into vllm-project:main Sep 16, 2024
72 checks passed

ElizaWszola mentioned this pull request Sep 18, 2024

Integrate fused Mixtral MoE with Marlin kernels #7079

Closed

Jeffwan pushed a commit to aibrix/vllm that referenced this pull request Sep 19, 2024

[Kernel] Enable 8-bit weights in Fused Marlin MoE (vllm-project#8032)

4678370

Co-authored-by: Dipika <dipikasikka1@gmail.com>

dtrifiro pushed a commit to opendatahub-io/vllm that referenced this pull request Sep 27, 2024

[Kernel] Enable 8-bit weights in Fused Marlin MoE (vllm-project#8032)

1965aad

Co-authored-by: Dipika <dipikasikka1@gmail.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Kernel] Enable 8-bit weights in Fused Marlin MoE #8032

[Kernel] Enable 8-bit weights in Fused Marlin MoE #8032

ElizaWszola commented Aug 30, 2024 •

edited

Loading

github-actions bot commented Aug 30, 2024

dsikka left a comment

dsikka Aug 30, 2024

dsikka commented Aug 30, 2024

dsikka commented Aug 30, 2024

dsikka Aug 30, 2024

dsikka Aug 30, 2024 •

edited

Loading

dsikka commented Aug 30, 2024

dsikka commented Sep 4, 2024

dsikka left a comment

dsikka Sep 13, 2024

ElizaWszola Sep 13, 2024

mgoin left a comment

[Kernel] Enable 8-bit weights in Fused Marlin MoE #8032

[Kernel] Enable 8-bit weights in Fused Marlin MoE #8032

Conversation

ElizaWszola commented Aug 30, 2024 • edited Loading

github-actions bot commented Aug 30, 2024

dsikka left a comment

Choose a reason for hiding this comment

dsikka Aug 30, 2024

Choose a reason for hiding this comment

dsikka commented Aug 30, 2024

dsikka commented Aug 30, 2024

dsikka Aug 30, 2024

Choose a reason for hiding this comment

dsikka Aug 30, 2024 • edited Loading

Choose a reason for hiding this comment

dsikka commented Aug 30, 2024

dsikka commented Sep 4, 2024

dsikka left a comment

Choose a reason for hiding this comment

dsikka Sep 13, 2024

Choose a reason for hiding this comment

ElizaWszola Sep 13, 2024

Choose a reason for hiding this comment

mgoin left a comment

Choose a reason for hiding this comment

ElizaWszola commented Aug 30, 2024 •

edited

Loading

dsikka Aug 30, 2024 •

edited

Loading