All pull requests

Pull requests

• [WS1][Qwen-Image][CUDA] Add stride-aware multi-axis RoPE with compile support

  #623·dandan20031031 opened Oct 28, 2026·

  ·
  1

• [ROCm] Add strict provider manifest core and GDN candidate observations

  #622·Anoymoguy opened Oct 28, 2026·

  ·
  1

• feat(adaln): add indexed CUDA/Triton modulation and explicit strict qualification

  #621·hongleng opened Oct 28, 2026Contributor·

  ·
  35

• Feat/dsv4 issue 58 ep integrity

  #620·hihaluemen opened Oct 28, 2026Contributor·

  ·
  17

• # [DSV4][ROCm][P6] Add reusable cross-config first-error diagnostics

  deepseek-P6DSv4platform: rocm
  #619·hihaluemen opened Oct 27, 2026Contributor·

  ·
  11

• [DSV4][ROCm] Add contract-bound C4 KV compressor (task 19)

  deepseek-P2DSv4platform: rocm
  #618·inaniloquentee opened Oct 27, 2026Collaborator·

  ·
  6

• [ROCm][DSV4][P6] Implement fixed-order local input-gradient combine (task 56)

  deepseek-P6DSv4platform: rocm
  #617·hihaluemen opened Oct 27, 2026Contributor·

  ·
  12

• [DSV4][ROCm] Add contract-bound C128 KV compressor (task 18)

  deepseek-P2DSv4platform: rocm
  #616·inaniloquentee opened Oct 27, 2026Collaborator·

  ·
  8

• [ROCm][DSV4][P6] Implement fixed-order output combine (tasks 53-55)

  deepseek-P6DSv4platform: rocm
  #615·Flink-ddd opened Oct 27, 2026Collaborator·

  ·
  7

• [ROCm][DSV4][P5] Add fixed-order route-weight gradient (task 49)

  deepseek-P5DSv4platform: rocm
  #614·Flink-ddd opened Oct 27, 2026Collaborator·

  ·
  5

• [ROCm][DSV4][P5] Integrate expert-module training (task 48)

  deepseek-P5DSv4platform: rocm
  #613·Flink-ddd opened Oct 27, 2026Collaborator·

  ·
  7

• [ROCm][DSV4][P6] Implement expert result restoration to original tokens (task 52)

  deepseek-P6DSv4platform: rocm
  #612·hihaluemen opened Oct 27, 2026Contributor·

  ·
  16

• [ROCm][DSV4][P6] Implement output combine interface and test preparation (task 51)

  deepseek-P6DSv4platform: rocm
  #611·hihaluemen opened Oct 27, 2026Contributor·

  ·
  10

• [ws1][Muse-Glimmer-30B] SwiGLU: FP64 references, Muse shapes and backend parity

  Muse-GlimmerWS1
  #610·sevleete opened Oct 27, 2026·

  ·
  10

• [DSV4][ROCm][P3] Add deterministic router auxiliary loss (task 34)

  deepseek-P3DSv4platform: rocm
  #609·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  7

• [DSV4][ROCm][P3] Add router backward and result records (task 33)

  deepseek-P3DSv4platform: rocm
  #608·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  10

• [ROCm][DSV4][P5] Implement clamp_swiglu_weighted (task 46)

  deepseek-P5DSv4platform: rocm
  #607·Flink-ddd opened Oct 27, 2026Collaborator·

  ·
  6

• [DSV4][ROCm][P3] Add learned-router weights and deterministic capacity (task 32)

  deepseek-P3DSv4platform: rocm
  #606·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  3

• [DSV4][ROCm][P3] Add stable Top-6 by score and global expert ID (task 31)

  deepseek-P3DSv4platform: rocm
  #605·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  5

• [DSV4][ROCm][P3] Add BI FP32 Hash routing with versioned tables (task 30)

  deepseek-P3DSv4platform: rocm
  #604·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  6

• [DSV4][ROCm][P3] Add BI FP32 router sqrt-softplus and stable gradients (task 29)

  deepseek-P3DSv4platform: rocm
  #603·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  7

• [DSV4][ROCm][P3] Add BI FP32 router Gate projection with upstream GEMM (task 28)

  deepseek-P3DSv4platform: rocm
  #602·maxiaosong1124 opened Oct 27, 2026Collaborator·

  ·
  5

• [ROCm][DSV4][P5] Add shared BF16 LoRA autograd (task 45)

  deepseek-P5DSv4platform: rocm
  #601·Flink-ddd opened Oct 27, 2026Collaborator·

  ·
  5

• [DSV4][ROCm] Add BI quantized Indexer ICV with upstream kernels

  deepseek-P2DSv4platform: rocm
  #599·inaniloquentee opened Oct 26, 2026Collaborator·

  ·
  11

• feat(activation): add shared parameterized final logit softcap

  #597·sevleete opened Oct 26, 2026·

  ·
  1

  19