All pull requests

Pull requests

• train 1114021234: #2885 #2722 #2961 (CI only)

  #3290·ashhart opened Nov 14, 2026Owner·

  ·

• dsv41 metal: vector greedy picks and kept sampling scratch, same picks

  #3288·ashhart opened Nov 14, 2026Owner·

  ·

• dsv41 metal: FP8 decode rows load first and meet tile roots in one simdgroup, same bits

  #3287·ashhart opened Nov 14, 2026Owner·

  ·

• train 1114020153: #2767 #2768 (CI only)

  #3286·ashhart opened Nov 14, 2026Owner·

  ·

• server: opt-in committed token IDs on chat and completion streams

  #3285·ashhart opened Nov 14, 2026Owner·

  ·
  1

• Prefill 8bit prompt windows

  #3280·scchumbletech opened Nov 14, 2026·8 tasks done·

  ·

• DeepSeek-V4.1 on Metal: opt-in shared context pool, TF_DSV41_POOL_TOKENS (stacked on #2983)

  #3278·gilby opened Nov 13, 2026Contributor·8 tasks done·

  ·

• flash next metal: TF_FLASHNEXT_SCALE_INDEX on M5 too

  #3277·ashhart opened Nov 13, 2026Owner·

  ·

• CI: drop the nightly Linux host shard; test-host runs all of it on every push

  #3276·ashhart opened Nov 13, 2026Owner·

  ·

• perf(hip): on gfx1151 a plain prompt of 1 or 2 rows takes the 2-row K-parallel tile

  #3273·AbrahamCete opened Nov 13, 2026Contributor·8 tasks done·

  ·
  1

  1

• Flash Next on M5: kept DeltaNet states and half nibbles, 6.5% faster decode, same tokens

  #3271·ashhart opened Nov 13, 2026Owner·

  ·

• metal: Spark2.5's and Llama's prompt attention from one source, Llama at WINDOW and RING 0 (#2251)

  #3269·jschmied opened Nov 13, 2026Contributor·1 task done·

  ·
  2

• info prices a Qwen3.8-27B-class Metal start from the charge serve makes (#2761)

  #3268·cshintov opened Nov 13, 2026Contributor·

  ·
  2

• server: the next turn's background prefill carries the reply's reasoning

  #3264·ashhart opened Nov 13, 2026Owner·

  ·

• Bench: a served A/B client for prompt-cache eviction on agent traffic (#3241)

  #3263·ashhart opened Nov 13, 2026Owner·

  ·

• Prompt cache: TF_CACHE_EVICT=decayed_resume, up to 34% fewer rebuilt rows on agent traces (opt-in)

  #3258·ashhart opened Nov 13, 2026Owner·

  ·

• qwen3_5 cuda: a dense MLP's NVFP4 prompt gate, up and act in one lean launch, the same bytes

  #3252·jschmied opened Nov 13, 2026Contributor·

  ·

• Nemotron on HIP: read the MTP head file at its own width (part 3 of 3)

  #3251·gilby opened Nov 13, 2026Contributor·7 of 8 tasks·

  ·
  1

  1

• glm cuda: each decode window width keeps its faster expert form

  #3249·Bizuayeu opened Nov 13, 2026Contributor·8 tasks done·

  ·

• CLI: tensorfold launch <agent> runs Claude Code, Codex, OpenCode or Hermes against the local server

  #3248·ashhart opened Nov 13, 2026Owner·

  ·
  1

• Prompt cache replay: agent-style traces through the real store (#3241, measurement)

  #3247·ashhart opened Nov 13, 2026Owner·

  ·

• Images: stream step progress and previews while a picture generates

  #3245·ashhart opened Nov 13, 2026Owner·

  ·
  1

  1

• Images: n from 1 to 4 a request, each picture with its own reported seed

  #3244·ashhart opened Nov 13, 2026Owner·

  ·
  1

• Nemotron on HIP: read each matrix at its own oQ width (part 2 of 3)

  #3242·gilby opened Nov 13, 2026Contributor·8 tasks done·

  ·
  2

• qwen27: Qwen3.6-35B-A3B on Metal on the routed path, with staged references against float64

  #3239·cshintov opened Nov 13, 2026Contributor·

  ·