karylab_agents/vllm

Fork 0

forked from Karylab-cklius/vllm

[NVIDIA] Fix DeepSeek V4 mHC warmup coverage #1

Open

karylab_agents wants to merge 1 commits from fix/deepseek-v4-nvidia-mhc-warmup into main

merge into: :main

:main

:ci/gitea-pr-search-compat

:ci/build-gb10-arm64-image

:fix/deepseek-v4-nvidia-mhc-warmup

:kimi-k3

:wentao-remove-multiple-dead-codes

:wentao-optimize-workspace-reuse

:codex/fix-cuda12-deepep-nccl

:k3-dspark-ar-fusion

:codex/test-ci-retry-failure

:dependabot/pip/minor-update-849870308e

:dependabot/github_actions/astral-sh/setup-uv-8.3.2

:dependabot/github_actions/astral-sh/setup-uv-8.2.0

:dependabot/github_actions/astral-sh/setup-uv-8.3.0

:wentao-fix-mypy-models-a-b

:wentao-support-rms-norm-uncontiguous

:bz/rust-frontend-reattach

:k3-release

:releases/v0.26.0

:bz/attractive-termite

:wentao-epd-support-for-MRv2

:wentao-fix-ci-pre-commit

:mgoin/fix-piecewise-capture-attention

:bz/yeasty-cephalopod

:bz/rust-code-coverage

:agent/amd-hf-offline-retry

:extensible-kv-cache

:lwilkinson/kv-layout/core-standardize

:cuda-arch-fixup

:zen5-image-build

:wentao-mrv2-enable-all-moe

:bz/fresh-hippopotamus

:tml-inkling

:agent/symm-mem-nvshmem

:dependabot/pip/protobuf-7.35.1

:dependabot/pip/fsspec-2026.6.0

:wentao-fused-padding-to-kernel

:codex/layer-parallel-plan

:wentao-fix-es-v2-bug

:embed-norm-fusion

:releases/v0.25.1

:releases/v0.25.0

:longcat-2.0

:zhuohan/extensible-kv-cache-memory

:wentao-optimize-per-token-group-quant

:fix-gguf-test

:woosuk/gumbel-fp32-rand64

:woosuk/glm5-megamoe

:bz/human-ox

:worktree-fix-cudagraph-flaky

:bz/sleep-metric

:codex/fix-flashinfer-fp8-moe-api-compat

:codex/fix-deepseek-v4-flashmla-cache-shape

:glm5--fp32-router

:dsv4-routing-from-trttllm

:woosuk/mrv2-draft-gumbel-decouple

:wentao-wq_b-to-cudagraph

:bz/civil-mouse

:amd-collab-iter0

:woosuk/dsv4-sp

:dependabot/github_actions/actions/setup-python-6.3.0

:bz/tolerate-oov-prompt

:codex/triton-warmup-compile-keys

:claude/pedantic-merkle-d957af

:woosuk/glm5-gemm

:releases/v0.24.0

:manual-act-quant-fusion-llama

:minimax-m3-perf

:woosuk/triton-fix

:wentao-fix-v2-mrope

:worktree-fix-ci-export-subshell

:bz/minimax-m3-engine-parser

:bz/logprobs-newtype

:speed-up-sp-tests-v2

:speed-up-sp-tests

:worktree-agent-ae80686f16ccbc350

:codex/rocm-artifact-tensorizer

:perf/push-allreduce-2buffer

:pr-44891

:releases/v0.23.0

:mnnvl_kv_transfer

:bz/refactor-build-rust-setup

:bz/bridge-rust-tool-parser-to-py

:releases/v0.22.1

:ucx_oneshot_ar

:dependabot/pip/pyrate-limiter-4.1.0

:dependabot/pip/fsspec-2026.4.0

:dependabot/github_actions/actions/setup-python-6.2.0

:as-of-2026-06-02

:luka/vllm-ir/compile-op

:revert-40687-matthias.skinny-gemm-n5

:releases/v0.22.0

:coverage-test-cosmos3

:coverage-test-kv-triton

:coverage-test-jais

:coverage-test-tilelang

:dependabot/pip/quack-kernels-gte-0.4.1

:worktree-coverage-test-mapping

:wentao-fix-flashinfer-layout

:wentao-fix-v2-test_spec_decode_acceptance_length

:wentao-deprecate-embed&token_classify

:dependabot/pip/protobuf-7.34.1

:ci/h200-35gb-remaining-20

:wentao-optimize-model-runner-v2-sampler

:migrate-safe-jobs-to-h200-mig

:releases/v0.21.0

:upgrade-cutedsl

:fix-mig-nvml-workaround

:migrate-gpu1-to-h200-18gb-remaining

:as-of-2026-05-12

:worktree-migrate-gpu1-to-h200-18gb-mig

:feat/tokenspeed_mla_upstream

:fix/topk-hash-indices-dtype

:ci/narrow-basic-correctness-deps

:ci/narrow-models-language-deps

:ci/narrow-models-multimodal-deps

:ci/narrow-models-basic-deps

:ci/narrow-entrypoints-deps

:wentao-model-runner-v2-support-stock-torch-compile

:releases/v0.20.2

:wentao-optimize-dcp-and-add-comm-func

:tokenspeed

:gemma4-mtp

:luka/vllm-ir-nits

:chinmay-amd-snapshot

:khluu/trigger-perf-eval-nightly

:wentao-optimize-pooling-forward

:releases/v0.20.1

:deepep-v2-integration

:releases/v0.20.1-python-from-source

:khluu/release-registry-cache

:khluu/vllm-base-uv-python

:khluu/release-v0.20.1-uv-python

:wentao-remove-dead-code

:khluu/b200-k8s-job-fixes

:khluu/b200-k8s-ci-smoke-20260429

:copilot/update-test-conftest-to-use-moe-backend

:copilot/add-sp-min-token-to-e2e-tests

:dsv4-pd-fixes

:fix-nixl-dockerfile

:deprecate-timeout

:releases/v0.20.0

:woosuk/fast-topk

:claude/slack-session-JTjDk

:luka/vllm-ir/triton

:lora-test

:wentao-optimize-pooling-by-ragged-tensor

:wentao-fix-ci-destroy

:wentao-update-batch-invariant-docstring

:luka/vllm-ir/rms-norm-batch-invariant

:bugfix/37931-nvfp4-batched-all2all

:codex/37931-flashinfer-cutedsl-batched-one-sided

:releases/v0.19.1

:cutlass_fa3_mla_sparse

:rebase-fa3-mla-sparse

:woosuk/ds-exp-2

:indexer_multistream

:khluu/mig

:woosuk/ds-exp-ag

:redhat-h100-testing

:ci/reorder-release-pipeline

:disable-image-build-per-commit

:claude/zen-banach

:fix/flashinfer-nvfp4-cross-row-scale-corruption

:khluu/gemma3

:wentao-cache-is_sleep

:wentao-fix-v2-is_prefiliing

:upgrade-transformers-compressed-tensors

:khluu/0190-540

:khluu/b200_k8s

:khluu/group_commands

:khluu/build0405

:revert-batch-kv-cache-swap-38460

:wentao-skip-work-when-empty

:releases/v0.19.0

:khluu/gemma2

:khluu/rocm_gemma

:compile-only-pr1

:khluu/automate-release-dockerhub-push

:releases/v0.18.1

:tms/fix-nan

:wentao-optimize-async-scheduling-copy

:wentao-fix-ci-batch-invariant-issue

:khluu/mig-small-model-swaps

:cursor/test-quality-improvements-eeea

:sm103

:vadim/qwen35-no-deppgemm

:woosuk/mrv2-expert-indices

:kernel-block-size-alignment-ssm

:fix_nixl_get_finished_handshake_failure

:remove-fp4-moe-env-var-clean

:releases/v0.18.0

:fix_nixl_triton_attn

:prometheus-cudagraph-pct

:khluu/cherrypick37322

:claude/optimize-weight-loading-7FlLd

:gb200-0317

:claude/nervous-meitner

:mrv2-ci-test

:tms/nvfp4-nan-contamination-test

:zhuohan/remove-unnecessary-instance_id-setup

:woosuk/mrv2-cudagraph-attn-fix

:releases/v0.17.1

:luka/fix-rms-quant-non-contiguous

:cursor/main-branch-failure-triage-f8d5

:cursor/VLLM-94-usage-stats-v2-design-584f

:vllm-dashboard

:fix/eplb-nvfp4-modelopt

:fix/eplb-prometheus-metrics

:fix/eplb-debug-logging

:fix/eplb-balancedness-metric

:releases/v0.17.0

:openai226

:0.17.0take2

:wentao-fix-dcp-IMA-for-v2

:wentao-fix-amd-ci-test-others-bug

:wentao-optimize-model-runner-v2-prepare_inputs

:pcp-alt

:releases/v0.16.0

:fix-mtp-dummy-run-assertion

:v0.16.0-torch291

:amd_dev

:zhuohan/redundant-pooling-check

:v0.16.0-before210

:v0.16.0-cu128

:fix_fi_cutlass

:add-cuda-12.8-wheel

:qwen3_5_fp8

:builder-nvcc-toolchain

:khluu/2/releases/v0.16.0

:khluu/releases/v0.16.0

:builder-cuda-version

:wentao-dcp-support-for-v2

:fix_moe_test_flashinfer

:release

:dockerignore_deps

:khluu/glm5

:fix-mtp

:khluu/feb11

:khluu/h200

:releases/v0.15.0

:khluu/disable_h200_x8

:bump_numba

:feat-k2.5-support

:releases/v0.14.1

:khluu-patch-1

:integrate_aiter_batched_deepgemm

:rocm_silu_mul_quant

:wentao-enable-flashinfer-moe-fp4-by-default

:releases/v0.14.0

:wentao-prefer-sysmem-comm

:andy-neuma-ibm-smoke

:khluu/test_ami

:wentao-fix-torch-compile-issue

:releases/v0.13.0

:wentao-update-torch-to-2.9.1

:ghsa-mcmc-2m55-j8jj

:wentao-fix-qwen3vl-launch-bug

:releases/v0.12.0

:batched_triton_fallback

:zhuohan/remove-redundant-argument

:wentao-fix-python-install-ci-error

:deepseek_optimizations_alex_rob

:skip-lmfe-tests

:releases/v0.11.2

:releases/v0.11.1

:rebased_fi_moe

:revert-27600-torch-utils-import

:zhuohan/revert-26709

:revert-26740-wentao-optimize-startup-log-2

:woosuk/test-router

:zhuohan/moe-kernel-experiment

:zhuohan/remove-virtual-engine

:woosuk/router-nixl

:update_from_kv_xfer_finished_race_fix

:releases/v0.11.0

:releases/v0.10.2

:codex/remove-vllm-v0-engine-references-from-docs

:wye-refactor-w8a8-quant

:simon-mo-patch-1

:fix_ds_eagle

:maybe_fix_hang_2

:dbo-cudagraph-size-cherry

:support_global_dp_logging

:codex/remove-raydistributedexecutor-from-v0-engine

:amd_mori

:split_kv_cache_init

:copilot/fix-31e676e9-a4af-4ed2-b74d-19d27f0a57b2

:copilot/fix-cudagraph-flag-combination

:il_tool

:copilot/fix-870996da-9146-438e-9a52-cdc6c1743086

:copilot/fix-584be906-f283-4e17-8776-c14111357ee7

:copilot/fix-56244f30-e76a-41ed-beaf-3bc9de22a2c9

:copilot/fix-c6914add-1b66-46d0-9948-c2e7b6f2259f

:releases/v0.10.1

:revert-22299-main

:remove_mamba_ssm

:acc-rate

:releases/v0.10.0

:wide_ep_working_branch_2

:wide_ep_working_branch

:revert-21550-chengji/fix-ci

:test-debug-lb

:tms/distributed_timeout

:7snzwi-codex/change-default-logging-behavior

:codex/change-default-logging-behavior

:mla_decode_any_head

:gpu_ids2

:gpu-ids

:fix-doc-build

:releases/v0.9.2

:topk_id_hack

:minus_x

:gemma3n-mm

:deep_full_cudagraph_fix

:deepep_tweaks

:fix-precommit

:releases/v0.9.1

:mergify/houseroad/config-update

:codex/add-pandas-and-datasets-to-requirements

:fp8_ep_dp

:releases/v0.9.0

:codex/update-arch-overview-md-with-vllm-v1-details

:benchmark_serving_test

:pil_image

:low_latency_opt

:woosuk-jf

:disable-sd

:v0.8.5

:pd_scheduling

:v0.8.4

:v1_fix_profiler

:fix_use_ep

:v0.8.3

:whisper-translate

:bench-latency

:v0.8.2

:v0.8.1

:v0.8.0

:mamba_tests

:running-deque

:bind_kv_caches

:reduce_scatter_comm

:amd-ci

:tpu_v1_optimized

:full_cudagraph

:mla_cuda_graphs

:qwen25vl

:tpu_v1

:moondream2

:torch_dynamo

:optimize-prefix-caching-scheduling

:fix-hashing-partial-blocks

:jax-tpu

Author SHA1 Message Date

Author	SHA1	Message	Date
karylab_agents	d789590bb6	Fix DeepSeek V4 NVIDIA mHC warmup coverage Expose the NVIDIA decoder mHC entry points to the warmup detector, warm the broadcast path, and compile the reachable split-key token sizes before serving. Backport of the fix from vllm-project/vllm#49707 (commits `32389a6b` and `21ef130e`).	2026-07-30 16:41:54 +08:00

karylab_agents

d789590bb6

Fix DeepSeek V4 NVIDIA mHC warmup coverage

Expose the NVIDIA decoder mHC entry points to the warmup detector, warm the broadcast path, and compile the reachable split-key token sizes before serving.

Backport of the fix from vllm-project/vllm#49707 (commits 32389a6b and 21ef130e).

2026-07-30 16:41:54 +08:00