[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 <[email protected]>

* [Kernel] Enable 8-bit weights in Fused Marlin MoE (vllm-project#8032) Co-authored-by: Dipika <[email protected]> * [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 <[email protected]> * [refactor] remove triton based sampler (vllm-project#8524) * [Frontend] Improve Nullable kv Arg Parsing (vllm-project#8525) Signed-off-by: Alex-Brooks <[email protected]> * [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 <[email protected]> * [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 <[email protected]> * [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 <[email protected]> * [doc] improve installation doc (vllm-project#8550) Co-authored-by: Andy Dai <[email protected]> * [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 <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> * [Core][Bugfix][Perf] Introduce `MQLLMEngine` to avoid `asyncio` OH (vllm-project#8157) Co-authored-by: Nick Hill <[email protected]> Co-authored-by: [email protected] <[email protected]> Co-authored-by: Robert Shaw <[email protected]> Co-authored-by: Simon Mo <[email protected]> * [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 <[email protected]> * [Model] Support Solar Model (vllm-project#8386) Co-authored-by: Michael Goin <[email protected]> * [AMD][ROCm]Quantization methods on ROCm; Fix _scaled_mm call (vllm-project#8380) Co-authored-by: Alexei-V-Ivanov-AMD <[email protected]> Co-authored-by: Michael Goin <[email protected]> * [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 <[email protected]> * [Kernel] Remove marlin moe templating on thread_m_blocks (vllm-project#8573) Co-authored-by: [email protected] * [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 <[email protected]> * [Hardware][AWS] update neuron to 2.20 (vllm-project#8676) Signed-off-by: omrishiv <[email protected]> * [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 <[email protected]> * [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 <[email protected]> * [MISC] rename CudaMemoryProfiler to DeviceMemoryProfiler (vllm-project#8703) * [Model][VLM] Add LLaVA-Onevision model support (vllm-project#8486) Co-authored-by: litianjian <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: DarkLight1337 <[email protected]> * [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 <[email protected]> * [Bugfix] fix docker build for xpu (vllm-project#8652) * [Core][Frontend] Support Passing Multimodal Processor Kwargs (vllm-project#8657) Signed-off-by: Alex-Brooks <[email protected]> * [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 <[email protected]> * [Kernel] (2/N) Machete - Integrate into CompressedTensorsWNA16 and GPTQMarlin (vllm-project#7701) Co-authored-by: mgoin <[email protected]> Co-authored-by: Divakar Verma <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> * [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 <[email protected]> * 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 <[email protected]> * 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 <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: DarkLight1337 <[email protected]> * [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 <[email protected]> * [Frontend] Batch inference for llm.chat() API (vllm-project#8648) Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Roger Wang <[email protected]> * [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 <[email protected]> * [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 <[email protected]> Signed-off-by: Alex-Brooks <[email protected]> Signed-off-by: Rui Qiao <[email protected]> Signed-off-by: Aaron Pham <[email protected]> Signed-off-by: Russell Bryant <[email protected]> Signed-off-by: omrishiv <[email protected]> Signed-off-by: Travis Johnson <[email protected]> Co-authored-by: ElizaWszola <[email protected]> Co-authored-by: Dipika <[email protected]> Co-authored-by: lewtun <[email protected]> Co-authored-by: Nick Hill <[email protected]> Co-authored-by: sasha0552 <[email protected]> Co-authored-by: Luka Govedič <[email protected]> Co-authored-by: youkaichao <[email protected]> Co-authored-by: Kevin Lin <[email protected]> Co-authored-by: Simon Mo <[email protected]> Co-authored-by: Joe Runde <[email protected]> Co-authored-by: Alex Brooks <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Rui Qiao <[email protected]> Co-authored-by: Isotr0py <[email protected]> Co-authored-by: sroy745 <[email protected]> Co-authored-by: chenqianfzh <[email protected]> Co-authored-by: Patrick von Platen <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Andy Dai <[email protected]> Co-authored-by: Alexey Kondratiev(AMD) <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> Co-authored-by: Daniele <[email protected]> Co-authored-by: Jiaxin Shan <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Aaron Pham <[email protected]> Co-authored-by: Alexander Matveev <[email protected]> Co-authored-by: [email protected] <[email protected]> Co-authored-by: Robert Shaw <[email protected]> Co-authored-by: afeldman-nm <[email protected]> Co-authored-by: Russell Bryant <[email protected]> Co-authored-by: Geun, Lim <[email protected]> Co-authored-by: Michael Goin <[email protected]> Co-authored-by: Gregory Shtrasberg <[email protected]> Co-authored-by: Alexei-V-Ivanov-AMD <[email protected]> Co-authored-by: Kuntai Du <[email protected]> Co-authored-by: Kunshang Ji <[email protected]> Co-authored-by: Charlie Fu <[email protected]> Co-authored-by: 盏一 <[email protected]> Co-authored-by: bnellnm <[email protected]> Co-authored-by: Amit Garg <[email protected]> Co-authored-by: William Lin <[email protected]> Co-authored-by: Niklas Muennighoff <[email protected]> Co-authored-by: saumya-saran <[email protected]> Co-authored-by: Pastel！ <[email protected]> Co-authored-by: omrishiv <[email protected]> Co-authored-by: zyddnys <[email protected]> Co-authored-by: youkaichao <[email protected]> Co-authored-by: rasmith <[email protected]> Co-authored-by: Divakar Verma <[email protected]> Co-authored-by: Huazhong Ji <[email protected]> Co-authored-by: litianjian <[email protected]> Co-authored-by: litianjian <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Lily Liu <[email protected]> Co-authored-by: Yuan <[email protected]> Co-authored-by: Yan Ma <[email protected]> Co-authored-by: Li, Jiang <[email protected]> Co-authored-by: Yanyi Liu <[email protected]> Co-authored-by: Jani Monoses <[email protected]> Co-authored-by: Lucas Wilkinson <[email protected]> Co-authored-by: Jee Jee Li <[email protected]> Co-authored-by: jiqing-feng <[email protected]> Co-authored-by: Hongxia Yang <[email protected]> Co-authored-by: Brendan Wong <[email protected]> Co-authored-by: Cody Yu <[email protected]> Co-authored-by: Peter Salas <[email protected]> Co-authored-by: Hanzhi Zhou <[email protected]> Co-authored-by: Andy <[email protected]> Co-authored-by: Travis Johnson <[email protected]> Co-authored-by: Archit Patke <[email protected]> Co-authored-by: zifeitong <[email protected]> Co-authored-by: sohamparikh <[email protected]>

Co-authored-by: Dipika <[email protected]>

[Bugfix] Streamed tool calls now more strictly follow OpenAI's format; ensures Vercel AI SDK compatibility (vllm-project#8272) [Frontend] Add progress reporting to run_batch.py (vllm-project#8060) Co-authored-by: Adam Lugowski <[email protected]> [Bugfix] Correct adapter usage for cohere and jamba (vllm-project#8292) [Misc] GPTQ Activation Ordering (vllm-project#8135) [Misc] Fused MoE Marlin support for GPTQ (vllm-project#8217) Add NVIDIA Meetup slides, announce AMD meetup, and add contact info (vllm-project#8319) [Bugfix] Fix missing `post_layernorm` in CLIP (vllm-project#8155) [CI/Build] enable ccache/scccache for HIP builds (vllm-project#8327) [Frontend] Clean up type annotations for mistral tokenizer (vllm-project#8314) [CI/Build] Enabling kernels tests for AMD, ignoring some of then that fail (vllm-project#8130) Fix ppc64le buildkite job (vllm-project#8309) [Spec Decode] Move ops.advance_step to flash attn advance_step (vllm-project#8224) [Misc] remove peft as dependency for prompt models (vllm-project#8162) [MISC] Keep chunked prefill enabled by default with long context when prefix caching is enabled (vllm-project#8342) [Bugfix] lookahead block table with cuda graph max capture (vllm-project#8340) [Bugfix] Ensure multistep lookahead allocation is compatible with cuda graph max capture (vllm-project#8340) [Core/Bugfix] pass VLLM_ATTENTION_BACKEND to ray workers (vllm-project#8172) [CI/Build][Kernel] Update CUTLASS to 3.5.1 tag (vllm-project#8043) [Misc] Skip loading extra bias for Qwen2-MOE GPTQ models (vllm-project#8329) [Bugfix] Fix InternVL2 vision embeddings process with pipeline parallel (vllm-project#8299) [Hardware][NV] Add support for ModelOpt static scaling checkpoints. (vllm-project#6112) [model] Support for Llava-Next-Video model (vllm-project#7559) Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> [Frontend] Create ErrorResponse instead of raising exceptions in run_batch (vllm-project#8347) [Model][VLM] Add Qwen2-VL model support (vllm-project#7905) Co-authored-by: Roger Wang <[email protected]> Co-authored-by: DarkLight1337 <[email protected]> [Hardware][Intel] Support compressed-tensor W8A8 for CPU backend (vllm-project#7257) [CI/Build] Excluding test_moe.py from AMD Kernels tests for investigation (vllm-project#8373) [Bugfix] Add missing attributes in mistral tokenizer (vllm-project#8364) [Kernel][Misc] register ops to prevent graph breaks (vllm-project#6917) Co-authored-by: Sage Moore <[email protected]> [Misc] Move device options to a single place (vllm-project#8322) [Speculative Decoding] Test refactor (vllm-project#8317) Co-authored-by: youkaichao <[email protected]> Pixtral (vllm-project#8377) Co-authored-by: Roger Wang <[email protected]> Bump version to v0.6.1 (vllm-project#8379) [MISC] Dump model runner inputs when crashing (vllm-project#8305) [misc] remove engine_use_ray (vllm-project#8126) [TPU] Use Ray for default distributed backend (vllm-project#8389) Fix the AMD weight loading tests (vllm-project#8390) [Bugfix]: Fix the logic for deciding if tool parsing is used (vllm-project#8366) [Gemma2] add bitsandbytes support for Gemma2 (vllm-project#8338) [Misc] Raise error when using encoder/decoder model with cpu backend (vllm-project#8355) [Misc] Use RoPE cache for MRoPE (vllm-project#8396) [torch.compile] hide slicing under custom op for inductor (vllm-project#8384) [Hotfix][VLM] Fixing max position embeddings for Pixtral (vllm-project#8399) [Bugfix] Fix InternVL2 inference with various num_patches (vllm-project#8375) Co-authored-by: DarkLight1337 <[email protected]> [Model] Support multiple images for qwen-vl (vllm-project#8247) Signed-off-by: Alex-Brooks <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: DarkLight1337 <[email protected]> [BugFix] lazy init _copy_stream to avoid torch init wrong gpu instance (vllm-project#8403) [BugFix] Fix Duplicate Assignment in Hermes2ProToolParser (vllm-project#8423) [Bugfix] Offline mode fix (vllm-project#8376) Signed-off-by: Joe Runde <[email protected]> [multi-step] add flashinfer backend (vllm-project#7928) [Core] Add engine option to return only deltas or final output (vllm-project#7381) [Bugfix] multi-step + flashinfer: ensure cuda graph compatible (vllm-project#8427) [Hotfix][Core][VLM] Disable chunked prefill by default and prefix caching for multimodal models (vllm-project#8425) [CI/Build] Disable multi-node test for InternVL2 (vllm-project#8428) [Hotfix][Pixtral] Fix multiple images bugs (vllm-project#8415) [Bugfix] Fix weight loading issue by rename variable. (vllm-project#8293) [Misc] Update Pixtral example (vllm-project#8431) [BugFix] fix group_topk (vllm-project#8430) [Core] Factor out input preprocessing to a separate class (vllm-project#7329) [Bugfix] Mapping physical device indices for e2e test utils (vllm-project#8290) [Bugfix] Bump fastapi and pydantic version (vllm-project#8435) [CI/Build] Update pixtral tests to use JSON (vllm-project#8436) [Bugfix] Fix async log stats (vllm-project#8417) [bugfix] torch profiler bug for single gpu with GPUExecutor (vllm-project#8354) bump version to v0.6.1.post1 (vllm-project#8440) [CI/Build] Enable InternVL2 PP test only on single node (vllm-project#8437) [doc] recommend pip instead of conda (vllm-project#8446) [Misc] Skip loading extra bias for Qwen2-VL GPTQ-Int8 (vllm-project#8442) [misc][ci] fix quant test (vllm-project#8449) [Installation] Gate FastAPI version for Python 3.8 (vllm-project#8456) [plugin][torch.compile] allow to add custom compile backend (vllm-project#8445) [CI/Build] Reorganize models tests (vllm-project#7820) [Doc] Add oneDNN installation to CPU backend documentation (vllm-project#8467) [HotFix] Fix final output truncation with stop string + streaming (vllm-project#8468) bump version to v0.6.1.post2 (vllm-project#8473) [Hardware][intel GPU] bump up ipex version to 2.3 (vllm-project#8365) Co-authored-by: Yan Ma <[email protected]> [Kernel][Hardware][Amd]Custom paged attention kernel for rocm (vllm-project#8310) [Model] support minicpm3 (vllm-project#8297) Co-authored-by: DarkLight1337 <[email protected]> [torch.compile] fix functionalization (vllm-project#8480) [torch.compile] add a flag to disable custom op (vllm-project#8488) [TPU] Implement multi-step scheduling (vllm-project#8489) [Bugfix][Model] Fix Python 3.8 compatibility in Pixtral model by updating type annotations (vllm-project#8490) [Bugfix][Kernel] Add `IQ1_M` quantization implementation to GGUF kernel (vllm-project#8357) [Kernel] Enable 8-bit weights in Fused Marlin MoE (vllm-project#8032) Co-authored-by: Dipika <[email protected]> [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 <[email protected]> [refactor] remove triton based sampler (vllm-project#8524) [Frontend] Improve Nullable kv Arg Parsing (vllm-project#8525) Signed-off-by: Alex-Brooks <[email protected]> [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 <[email protected]> [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 <[email protected]> [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 <[email protected]> [doc] improve installation doc (vllm-project#8550) Co-authored-by: Andy Dai <[email protected]> [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 <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> [Core][Bugfix][Perf] Introduce `MQLLMEngine` to avoid `asyncio` OH (vllm-project#8157) Co-authored-by: Nick Hill <[email protected]> Co-authored-by: [email protected] <[email protected]> Co-authored-by: Robert Shaw <[email protected]> Co-authored-by: Simon Mo <[email protected]> [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 <[email protected]> [Model] Support Solar Model (vllm-project#8386) Co-authored-by: Michael Goin <[email protected]> [AMD][ROCm]Quantization methods on ROCm; Fix _scaled_mm call (vllm-project#8380) Co-authored-by: Alexei-V-Ivanov-AMD <[email protected]> Co-authored-by: Michael Goin <[email protected]> [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 <[email protected]> [Kernel] Remove marlin moe templating on thread_m_blocks (vllm-project#8573) Co-authored-by: [email protected] [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 <[email protected]> [Hardware][AWS] update neuron to 2.20 (vllm-project#8676) Signed-off-by: omrishiv <[email protected]> [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 <[email protected]> [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 <[email protected]> [MISC] rename CudaMemoryProfiler to DeviceMemoryProfiler (vllm-project#8703) [Model][VLM] Add LLaVA-Onevision model support (vllm-project#8486) Co-authored-by: litianjian <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: DarkLight1337 <[email protected]> [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 <[email protected]> [Bugfix] fix docker build for xpu (vllm-project#8652) [Core][Frontend] Support Passing Multimodal Processor Kwargs (vllm-project#8657) Signed-off-by: Alex-Brooks <[email protected]> [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 <[email protected]> [Kernel] (2/N) Machete - Integrate into CompressedTensorsWNA16 and GPTQMarlin (vllm-project#7701) Co-authored-by: mgoin <[email protected]> Co-authored-by: Divakar Verma <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> [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 <[email protected]> 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 <[email protected]> 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 <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: DarkLight1337 <[email protected]> [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 <[email protected]> [Frontend] Batch inference for llm.chat() API (vllm-project#8648) Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Cyrus Leung <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Roger Wang <[email protected]> [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 <[email protected]> [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) [Frontend] OpenAI server: propagate usage accounting to FastAPI middleware layer (vllm-project#8672) [Bugfix] Ray 2.9.x doesn't expose available_resources_per_node (vllm-project#8767) Signed-off-by: darthhexx <[email protected]> [Misc] Fix minor typo in scheduler (vllm-project#8765) [CI/Build][Bugfix][Doc][ROCm] CI fix and doc update after ROCm 6.2 upgrade (vllm-project#8777) [Kernel] Fullgraph and opcheck tests (vllm-project#8479) [[Misc]] Add extra deps for openai server image (vllm-project#8792) [VLM][Bugfix] internvl with num_scheduler_steps > 1 (vllm-project#8614) rename PromptInputs and inputs with backward compatibility (vllm-project#8760) [Frontend] MQLLMEngine supports profiling. (vllm-project#8761) [Misc] Support FP8 MoE for compressed-tensors (vllm-project#8588) Revert "rename PromptInputs and inputs with backward compatibility (vllm-project#8760) (vllm-project#8810) [Model] Add support for the multi-modal Llama 3.2 model (vllm-project#8811) Co-authored-by: simon-mo <[email protected]> Co-authored-by: Chang Su <[email protected]> Co-authored-by: Simon Mo <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Roger Wang <[email protected]> [Doc] Update doc for Transformers 4.45 (vllm-project#8817) [Misc] Support quantization of MllamaForCausalLM (vllm-project#8822) [Misc] Update config loading for Qwen2-VL and remove Granite (vllm-project#8837) [Build/CI] Upgrade to gcc 10 in the base build Docker image (vllm-project#8814) [Docs] Add README to the build docker image (vllm-project#8825) [CI/Build] Fix missing ci dependencies (vllm-project#8834) [misc][installation] build from source without compilation (vllm-project#8818) [ci] Soft fail Entrypoints, Samplers, LoRA, Decoder-only VLM (vllm-project#8872) Signed-off-by: kevin <[email protected]> [Bugfix] Include encoder prompts len to non-stream api usage response (vllm-project#8861) [Misc] Change dummy profiling and BOS fallback warns to log once (vllm-project#8820) [Bugfix] Fix print_warning_once's line info (vllm-project#8867) fix validation: Only set tool_choice `auto` if at least one tool is provided (vllm-project#8568) [Bugfix] Fixup advance_step.cu warning (vllm-project#8815) [BugFix] Fix test breakages from transformers 4.45 upgrade (vllm-project#8829) [Installation] Allow lower versions of FastAPI to maintain Ray 2.9 compatibility (vllm-project#8764) [Feature] Add support for Llama 3.1 and 3.2 tool use (vllm-project#8343) Signed-off-by: Max de Bayser <[email protected]> [Core] rename`PromptInputs` and `inputs` (vllm-project#8876) [misc] fix collect env (vllm-project#8894) [MISC] Fix invalid escape sequence '\' (vllm-project#8830) Signed-off-by: Peter Pan <[email protected]> [Bugfix][VLM] Fix Fuyu batching inference with `max_num_seqs>1` (vllm-project#8892) [TPU] Update pallas.py to support trillium (vllm-project#8871) [torch.compile] use empty tensor instead of None for profiling (vllm-project#8875) [Kernel] AQ AZP 4/4: Integrate asymmetric quantization to linear method (vllm-project#7271) [Bugfix] fix for deepseek w4a16 (vllm-project#8906) Co-authored-by: mgoin <[email protected]> [Core] Multi-Step + Single Step Prefills via Chunked Prefill code path (vllm-project#8378) Co-authored-by: Varun Sundar Rabindranath <[email protected]> [misc][distributed] add VLLM_SKIP_P2P_CHECK flag (vllm-project#8911) [Core] Priority-based scheduling in async engine (vllm-project#8850) [misc] fix wheel name (vllm-project#8919) [Bugfix][Intel] Fix XPU Dockerfile Build (vllm-project#7824) Signed-off-by: tylertitsworth <[email protected]> Co-authored-by: youkaichao <[email protected]> [Misc] Remove vLLM patch of `BaichuanTokenizer` (vllm-project#8921) [Bugfix] Fix code for downloading models from modelscope (vllm-project#8443) [Bugfix] Fix PP for Multi-Step (vllm-project#8887) [CI/Build] Update models tests & examples (vllm-project#8874) Co-authored-by: Roger Wang <[email protected]> [Frontend] Make beam search emulator temperature modifiable (vllm-project#8928) Co-authored-by: Eduard Balzin <[email protected]> [Bugfix] Support testing prefill throughput with benchmark_serving.py --hf-output-len 1 (vllm-project#8891) [doc] organize installation doc and expose per-commit docker (vllm-project#8931) [Core] Improve choice of Python multiprocessing method (vllm-project#8823) Signed-off-by: Russell Bryant <[email protected]> Co-authored-by: youkaichao <[email protected]> [Bugfix] Block manager v2 with preemption and lookahead slots (vllm-project#8824) [Bugfix] Fix Marlin MoE act order when is_k_full == False (vllm-project#8741) Co-authored-by: Tyler Michael Smith <[email protected]> [CI/Build] Add test decorator for minimum GPU memory (vllm-project#8925) [Build/CI] Set FETCHCONTENT_BASE_DIR to one location for better caching (vllm-project#8930) [Model] Support Qwen2.5-Math-RM-72B (vllm-project#8896) [Model][LoRA]LoRA support added for MiniCPMV2.5 (vllm-project#7199) [BugFix] Fix seeded random sampling with encoder-decoder models (vllm-project#8870) Co-authored-by: Roger Wang <[email protected]> [Misc] Fix typo in BlockSpaceManagerV1 (vllm-project#8944) [Frontend] Added support for HF's new `continue_final_message` parameter (vllm-project#8942) [Kernel][Model] Varlen prefill + Prefill chunking support for mamba kernels and Jamba model (vllm-project#8533)

Co-authored-by: Dipika <[email protected]> Signed-off-by: Alvant <[email protected]>

Co-authored-by: Dipika <[email protected]> Signed-off-by: Amit Garg <[email protected]>

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 <[email protected]>

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 <[email protected]>

MengqingCao pushed a commit to MengqingCao/vllm that referenced this pull request Sep 30, 2024

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

558d305

Co-authored-by: Dipika <[email protected]>

siddharth9820 pushed a commit to axonn-ai/vllm that referenced this pull request Sep 30, 2024

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

4badf5f

Co-authored-by: Dipika <[email protected]>

Alvant pushed a commit to compressa-ai/vllm that referenced this pull request Oct 26, 2024

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

49247c0

Co-authored-by: Dipika <[email protected]> Signed-off-by: Alvant <[email protected]>

garg-amit pushed a commit to garg-amit/vllm that referenced this pull request Oct 28, 2024

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

bac09e3

Co-authored-by: Dipika <[email protected]> Signed-off-by: Amit Garg <[email protected]>

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