All pull requests

Pull requests

• perf(vl): optimize image process

  #5133·caikun-pjlab opened Nov 13, 2026Collaborator·

  ·

• fix(sampling): use int64 seeds so sampling works on Windows

  #5131·ZoerZhu opened Nov 12, 2026·

  ·

• fix(anthropic): honor disable_parallel_tool_use

  #5130·kodurd opened Nov 12, 2026Contributor·

  ·

• [Fix] Validate DeepSeek completion format under optimized Python

  #5129·tgghzyh opened Nov 11, 2026·

  ·

• [Fix] Mark TurboMind runtime dependencies as Linux-only

  #5128·tgghzyh opened Nov 11, 2026·

  ·
  1

• fix(api): report cached_tokens in /v1/completions usage

  #5124·wanfengcxz opened Nov 11, 2026Collaborator·

  ·
  2

• fix(api): apply thinking config in /v1/messages/count_tokens

  #5123·wanfengcxz opened Nov 11, 2026Collaborator·

  ·
  2

• [Fix] Require API key authentication for proxy node management routes

  #5122·uditsharma29 opened Nov 10, 2026·

  ·
  1

• [Feature] Support incremental routed-expert responses in the PyTorch backend

  #5121·RunningLeon opened Nov 10, 2026Collaborator·4 tasks done·

  ·
  2

• [Feature] Support weight-versioned Mooncake Store and routed-expert recovery for RL

  #5120·RunningLeon opened Nov 10, 2026Collaborator·2 of 4 tasks·

  ·

• fix: align DeepEP capacity for masked FP8 GEMM

  #5119·RunningLeon opened Nov 10, 2026Collaborator·3 of 4 tasks·

  ·
  2

• Stabilize CUDA completion timing test

  #5118·RunningLeon with CopilotAI opened Nov 10, 2026·

  ·
  1

  

• [Build] Upgrade flash-linear-attention to 0.6.0

  #5117·RunningLeon opened Nov 9, 2026Collaborator·3 of 4 tasks·

  ·

• [Fix] Avoid zero-token requests in random benchmarks

  #5116·emo-ocean opened Nov 9, 2026·

  ·

• fix(anthropic): count tokens for declared tools

  #5114·kodurd opened Nov 9, 2026Contributor·

  ·
  5

• fix(vl): honor request-local multimodal processor options

  #5113·RunningLeon opened Nov 9, 2026Collaborator·3 of 4 tasks·

  ·
  3

• fix(pytorch): use lora_alpha / sqrt(r) for rsLoRA adapters

  #5112·Andreas-Hum opened Nov 8, 2026·

  ·
  1

  1

• [Feature] Add opt-in GLM53 activation and W2 quantization contracts

  #5109·RunningLeon opened Nov 6, 2026Collaborator·3 of 4 tasks·

  ·

• [Feature] Optimize GLM-5.3 SM90 decode

  enhancement
  #5108·qescccczmr opened Nov 6, 2026Collaborator·5 of 6 tasks·

  ·

• fix(media): close HTTP sessions on every download exit

  #5106·silencelamb opened Nov 6, 2026Collaborator·4 tasks done·

  ·

• fix(vision): apply EXIF orientation before image preprocessing

  #5105·silencelamb opened Nov 6, 2026Collaborator·4 tasks done·

  ·

• fix(openai): preserve required response fields and optional usage counts

  #5104·silencelamb opened Nov 6, 2026Collaborator·4 tasks done·

  ·

• fix(openai): validate image detail and return standard errors

  #5103·silencelamb opened Nov 6, 2026Collaborator·4 tasks done·

  ·

• fix(glm5): honor per-request native image pixel budgets

  #5102·silencelamb opened Nov 6, 2026Collaborator·4 tasks done·

  ·

• fix(glm5): preserve developer instructions in chat templates

  #5101·silencelamb opened Nov 6, 2026Collaborator·4 tasks done·

  ·