All pull requests

Pull requests

• fix(qual): report unavailable code sandboxes early

  #2228·JiaxinD opened Nov 11, 2026Contributor·5 of 11 tasks·

  ·
  1

• fix(aiperf): do not fail TTS validity for silence the native model shares

  #2227·chaofengw-nv opened Nov 11, 2026Collaborator·3 of 11 tasks·

  ·
  1

  4

• fix(lance): run the native reference on PyTorch SDPA where flash-attn has no kernels

  #2223·chaofengw-nv opened Nov 11, 2026Collaborator·3 of 11 tasks·

  ·
  3

• fix(dinov3): name the failed feature criterion and its value

  #2222·chaofengw-nv opened Nov 11, 2026Collaborator·3 of 11 tasks·

  ·
  4

• fix(aiperf): install the wan22-ti2v-5b reference without build isolation

  #2221·chaofengw-nv opened Nov 11, 2026Collaborator·3 of 11 tasks·

  ·
  3

• perf(hstu): read the ranking record in one SAX pass

  #2218·MadFunMaker opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

• perf(hstu): skip the attention's causally masked key blocks

  #2217·MadFunMaker opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

• perf(hstu): multiply the attention mask in instead of selecting

  #2216·MadFunMaker opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

• feat(hstu): add native HSTU ranking family (recsys-examples KuaiRand-1K)

  #2215·MadFunMaker opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

• feat(minimax_h3): multi-GPU context-parallel DiT and tile-parallel VAE on TRT-RTX

  #2214·pkisfaludi-nv opened Nov 10, 2026Collaborator·

  ·
  1

• fix(aiperf): count InternLM2 prompts with the bundle's tokenizer.json

  #2209·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

  3

• perf(internvl): run the InternViT attention and MLP blocks in FP16

  #2208·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

  1

• perf(nemotron_h): prefill single-GPU prompts in multi-token chunks

  #2207·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  1

• fix(qual): share pinned native JSON tokenizers

  #2205·JiaxinD opened Nov 10, 2026Contributor·4 of 11 tasks·

  ·
  2

• fix(nemotron_speech_streaming): keep the FP8 attention projections in FP16

  #2204·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

  1

• perf(phi4_multimodal): prefill and decode over the cached prefix, not the whole KV allocation

  #2202·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  1

• fix(dinov3): run the ViT attention softmax in FP32

  #2200·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

  1

• fix(aiperf): give the qwen3-vl-8b reference the candidate's image budget

  #2198·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  3

• fix(fast_foundation_stereo): build kernel-4 stride-2 deconvolutions as four convolutions

  #2196·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

  1

• perf(k2_horizon_uno): prefill prompts in 64-row base-profile chunks

  #2194·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  3

• perf(glmasr): prefill the audio prompt in one launch of a prefill engine

  #2193·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  1

• perf(rwkv): prefill text prompts in multi-token chunks

  #2192·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  1

• perf(qwen3_5): prefill text prompts in multi-token chunks

  #2191·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  4

• fix(aiperf-qual): cut retrieval documents to what the bundle holds

  #2187·chaofengw-nv opened Nov 10, 2026Collaborator·3 of 11 tasks·

  ·
  1

  3

• perf(mamba): prefill text prompts in multi-token chunks

  #2184·chaofengw-nv opened Nov 10, 2026Collaborator·4 of 11 tasks·

  ·
  1

  1