All pull requests

Pull requests

• strix-hip train: the validated gfx1151 stack (build fix + suballoc + carveout + hipBLASLt)

  #197·AzeezIsh opened Oct 19, 2026Contributor·

  ·

• scheduler: tool-parameter markers from the served tokenizer (Nemotron tool calls cut at 1024 tokens)

  #196·AzeezIsh opened Oct 19, 2026Contributor·

  ·
  1

• qwen4_exp: serve NVIDIA's Qwen3.8-Flash-Next FP4-Local on its native formats (one GB10)

  #195·AzeezIsh opened Oct 18, 2026Contributor·7 of 9 tasks·

  ·
  3

• ci: devlab-strix validate + eval jobs

  #194·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• strix-hip: drop Nemotron's dead dequant_nvfp4_bf16 override (kernel-shadow check red on main since #169)

  #192·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• spark-storage: keepalive test waits on deadlines, not fixed sleeps (intermittent CI failure)

  #191·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• moe: HIP unit tests take the grouped dispatch they assert (6 #174 GLM tests failed or hung on HIP)

  #188·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• hc_token_fused: never on under HIP (strix-hip Flash-Next decode aborts on main)

  #187·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• hip shim: sub-allocate the sizes hipMalloc rounds up (Flash-Next fits under Linux; part of #145)

  #186·AzeezIsh opened Oct 18, 2026Contributor·

  ·
  1

• AMD train 2026-10-17: #166 (incl. #152, #153, #155) + #167 + #169 + #157, plus the HIP build fix

  #184·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• spark-runtime: don't report host MemAvailable as device free on a carve-out APU

  #182·AzeezIsh opened Oct 18, 2026Contributor·

  ·

• perf: strix-hip BF16 GEMMs via hipBLASLt (hip-p0→C5 train)

  needs-rebase
  #166·AzeezIsh opened Oct 6, 2026Contributor·

  ·
  3

• prefix cache: caching on = caching off at temperature 0 on SSM-hybrid models (fixes #158)

  #163·maplepaladin73 opened Oct 4, 2026Contributor·

  ·
  1

  32

• Qwen3.8-Flash-Next EXL3: Native CoopMK trellis MoE, device QSA, and 262K recipes (M5)

  needs-rebase
  #160·vcruz305 opened Oct 3, 2026·1 task done·

  ·
  5

• lm_head: one-pass BF16 batch2 LM head on every target (stacked on #153; GB10 Flash-Next MTP verify)

  #159·AzeezIsh opened Oct 2, 2026Contributor·

  ·
  2

• strix-hip: F16-operand prefill GEMM twins (exact NVFP4/FP8 dequant) for dense m128, routed MoE, w8a16

  needs-rebase
  #155·AzeezIsh opened Oct 2, 2026Contributor·

  ·
  5

• lm_head: one-pass BF16 batch2 LM head for K=2 verify on HIP (Flash-Next, −5.9 ms/verify step)

  #153·AzeezIsh opened Oct 2, 2026Contributor·

  ·
  3

• qsa: O(complete) radix-select top-k on the device by default (Flash-Next decode past the 2051-token bound)

  needs-rebase
  #152·AzeezIsh opened Oct 2, 2026Contributor·4 tasks·

  ·
  5

• strix-windows: Flash-Next launchers + prefix caching on by default (stacked on #80)

  #149·AzeezIsh opened Oct 1, 2026Contributor·4 tasks·

  ·
  10

• bench: retire the FP8 gate subjects -- NVFP4-only defaults on gb10

  #142·AzeezIsh opened Sep 30, 2026Contributor·6 tasks done·

  ·
  3

• ci: ratchet the number of distinct ATLAS_* env names (#117)

  #118·AzeezIsh opened Sep 27, 2026Contributor·

  ·
  1

• release: verified GB10 image 3fcf8ff record + one-command verify script (publish pending)

  #113·AzeezIsh opened Sep 27, 2026Contributor·

  ·
  2

• Qwen3.8-Flash-Next EXL3: MTP + native int8 decode for GDN/attention/lm_head (M6, follow-up to #91)

  needs-rebase
  #100·andrew-vogulkin opened Sep 26, 2026Contributor·4 of 6 tasks·

  ·
  11

• gemma-4-E2B integration: model loaders, vision/audio encoders, PLE, kernels

  needs-rebase
  #11·Marker689 opened Aug 26, 2026·

  ·
  4