All pull requests

Pull requests

• fix(qwen4_exp): reject fused GDN MTP decode for sigmoid-gated checkpoints

  #348·Timik232 opened Oct 22, 2026·

  ·
  1

  1

• feat(qwen4_exp): offload PLE ngram embedding to pinned host memory

  #346·Timik232 opened Oct 22, 2026·

  ·
  1

  1

• [BugFix] DSpark: load ignore-listed mtp wo_a/main_proj weights unquantized

  #344·Timik232 opened Oct 21, 2026·3 of 4 tasks·

  ·
  1

  1

• [BugFix] dsv4: initialize fp8 KV descale factors before fused insert

  #343·Timik232 opened Oct 21, 2026·3 of 4 tasks·

  ·
  1

  1

• [port] carry mx/v0.28.0-dev patches onto master (vLLM 0.29.0)

  #339·luffydream opened Sep 14, 2026Contributor·4 tasks·

  ·

• test: add kv offload tests aligned with upstream

  #334·lianjiezh opened Aug 10, 2026Collaborator·

  ·

• test: add some tests/kernels cases

  #333·lianjiezh opened Aug 10, 2026Collaborator·

  ·

• fix(envs) — correct bool parsing for MACA env vars + mctlass benchmark harness

  #332·LindseyMei opened Aug 9, 2026·

  ·

• feat(attn) — add autotuned triton decode attention tiles for MXC500

  #331·LindseyMei opened Aug 9, 2026·

  ·

• [moe] add more MXC500 fused-MoE tuned configs (DeepSeek-V2 / Mixtral-8x22B / DeepSeek-V3)

  #329·LindseyMei opened Aug 2, 2026·

  ·
  1

• Use max_seq_len instead of prefill_seq_lens.max().item() to reduce CPU-side overhead

  #325·liunianxuxie opened Jul 21, 2026·

  ·
  2

• Add tuned MoE Triton configs for MetaX C500 (MXC500)

  #323·LindseyMei opened Jul 13, 2026·4 tasks done·

  ·

• [CI] run pre-commit on the default branch

  #322·Lfan-ke opened Jul 13, 2026·

  ·
  1

• [BugFix] route awq_dequantize through the upstream custom op

  #321·Lfan-ke opened Jul 13, 2026·

  ·
  1

• test: port repetition detection tests from upstream vLLM

  #317·zhouxj52 opened Jul 6, 2026·

  ·
  1

• docs: add Qwen/Qwen2.5-1.5B-Instruct to supported models

  #312·LindseyMei opened Jul 1, 2026·

  ·

• Add vLLM-metax vllm metax server log summary

  #311·ghangz opened Jul 1, 2026·

  ·
  5

• Fix MetaX vLLM 0.23 MTP compatibility

  #294·xzh25 opened Jun 22, 2026·

  ·
  11

• docs: align vllm source branch with 0.23

  #293·xzh25 opened Jun 21, 2026·

  ·
  2

• tests: sync v1 kv cache tests for vllm 0.23

  #292·xzh25 opened Jun 21, 2026·

  ·
  16

• [FA] reuse max_seq_len for prefill_max_seq_len

  #286·sxvvv opened Jun 13, 2026·

  ·

• [Doc] document pip wheel install path and version matching

  #285·Zymonody7 opened Jun 10, 2026·4 tasks done·

  ·
  1

• [BugFix] correct topk_softmax/topk_sigmoid routing weights at (128, 8)

  #284·Zymonody7 opened Jun 10, 2026·3 of 4 tasks·

  ·

• [BugFix] call ops.top_k_per_row_prefill in int8 indexer prefill

  #283·Zymonody7 opened Jun 10, 2026·3 of 4 tasks·

  ·

• test: add real-GPU numerical tests for MoE topk_softmax/topk_sigmoid gating

  #269·sxvvv opened Jun 6, 2026·

  ·
  1