All pull requests

Pull requests

• build: the portable-names test runs every time, since a reused pass can't see a name added since

  #2311·ashhart opened Nov 6, 2026Owner·1 task·

  ·

• Lean: 30 more comments reworded to one line in core and the Metal kernels

  #2309·ashhart opened Nov 6, 2026Owner·

  ·

• Contract: masked non-finite values across every Metal prompt attention, the findings registered as known (refs #1009)

  #2308·ashhart opened Nov 6, 2026Owner·

  ·

• profile: register the pool-page switches for Dory, Gemma 4, Spark2.5 and Nemotron's Metal paging

  #2306·devopsfortunato opened Nov 6, 2026Contributor·4 tasks done·

  ·
  1

• Tests: pins for the documented limits, defaults and flag dependencies, a replay of the prompt cache's eviction, and a timing of its index

  #2305·ashhart opened Nov 6, 2026Owner·

  ·

• CUDA: a request's stats list its prompt pass widths for the 27B, GLM and Nemotron (stacked on #2293)

  #2304·ashhart opened Nov 6, 2026Owner·

  ·

• Tests: the 27B decode control edits #2106's masked line, so M5 contract runs finish

  #2303·ashhart opened Nov 6, 2026Owner·

  ·
  1

  1

• ricursive cuda: Ripple-65B-Preview NVFP4 on the native CUDA engine

  #2302·jrb-ai opened Nov 6, 2026Contributor·

  ·
  1

  1

• Docs: api.md, README and vision.md match what the server reads and writes

  #2299·ashhart opened Nov 6, 2026Owner·

  ·

• qwen27: a prompt frame below 128 rows is refused by name, not as TargetBinding

  #2298·ashhart opened Nov 6, 2026Owner·

  ·

• docs: what a no-thinking GLM-5.3 prompt is on each engine (#1911)

  #2297·devopsfortunato opened Nov 6, 2026Contributor·4 tasks done·

  ·
  1

• Docs: Flash Next CUDA's TF_FLASHNEXT_CONFIDENCE defaults to -0.4, not 0.60

  #2296·ashhart opened Nov 6, 2026Owner·

  ·

• Lean: 73 more comments reworded to one line in kernels, tests, cluster and fabric

  #2295·ashhart opened Nov 6, 2026Owner·

  ·

• moss_image: the transformer gated block by block against float32, on Metal and CUDA

  #2294·BobClawblaw opened Nov 6, 2026Contributor·

  ·
  2

• Metal: a request's stats list its prompt pass widths, on every text family

  #2293·ashhart opened Nov 6, 2026Owner·

  ·

• Lean: 100 findings fixed in core, server, cluster and the families, with no change in behaviour

  #2292·ashhart opened Nov 6, 2026Owner·

  ·

• Tests: the MOSS-Image teacher gate compares each step with the teacher, not only with a repeat (#1529)

  #2291·ashhart opened Nov 5, 2026Owner·

  ·

• build: a -Dgpu server's modules import the profile registry

  #2290·atakann opened Nov 5, 2026Contributor·9 tasks done·

  ·
  1

• CUDA: every synchronous readback stages on integrated GPUs, as download does (#2102)

  #2289·ashhart opened Nov 5, 2026Owner·

  ·

• Persistence: the key gaps #2163 leaves, for every family (closes #2131)

  #2287·ashhart opened Nov 5, 2026Owner·

  ·

• qwen3 cuda: Qwen3 embedding checkpoints (BF16) on the native CUDA engine (#2261, 1 of 2)

  #2285·Arminova opened Nov 5, 2026Contributor·9 tasks done·

  ·
  1

• cli: run the pull tests' fake hub on std.Io.net so Windows builds them

  #2284·ElderOrb opened Nov 5, 2026·4 tasks done·

  ·
  4

• qwen27 metal: keys and values in core pool pages behind TF_QWEN27_METAL_KV_PAGES=1 (#1293)

  #2283·jschmied opened Nov 5, 2026Contributor·

  ·
  1

• core: Gemma 4 and Qwen 3.5 on CUDA load with direct reads, 28-31 % faster cold; Flash Next shares the read-ahead

  #2282·jschmied opened Nov 5, 2026Contributor·

  ·
  2

• moss image: serve pictures on CUDA (encoder, decoder and host for #1573)

  #2281·devopsfortunato opened Nov 5, 2026Contributor·5 tasks done·

  ·
  4