All pull requests

Pull requests

• lfm2_moe CUDA: attention pages from the shared KV pool, and the prompt cache (stacked on #1492)

  #1503·jschmied opened Oct 27, 2026Contributor·

  ·

• lfm3: add native Metal text and vision support for MLX 4/6/8-bit

  #1502·BlivionIaG opened Oct 27, 2026Contributor·5 of 9 tasks·

  ·
  1

• KV pool: host tier for the core pool (5/8)

  #1501·BlivionIaG opened Oct 27, 2026Contributor·

  ·

• exl3: TF_EXL3_SPLITK holds the grouped experts' K split for an A/B

  #1500·BadAd84 opened Oct 27, 2026Contributor·

  ·

• lanes: the first-window tree reads the Model instead of a constant

  #1499·BadAd84 opened Oct 27, 2026Contributor·

  ·

• glm: the family's own prompt-lookup copy drafts behind the core Proposers hook

  #1498·BadAd84 opened Oct 27, 2026Contributor·

  ·

• glm: size the DFlash2 raw quantization buffer for every world

  #1496·BadAd84 opened Oct 27, 2026Contributor·

  ·

• roce: serve worlds two to four in the one-shot gather, pairing rails by shared subnet

  #1495·BadAd84 opened Oct 27, 2026Contributor·

  ·

• lfm2_moe: serve LFM2.5-24B-A2B on CUDA (plain serve, shared kernels; stacked on #1377)

  #1492·jschmied opened Oct 27, 2026Contributor·

  ·

• tools: authenticate cache benchmark probes with explicit bearer keys

  #1483·Maxrio9 opened Oct 27, 2026Contributor·

  ·
  1

  1

• qwen3_5 cuda: require long_match (20 tokens) for the 27B's root copies

  #1477·diwakergupta opened Oct 27, 2026·9 tasks done·

  ·

• qwen3_5 cuda: host-mapped EXL3 embeddings and shared mixer/MLP scratch (stacked on #1474)

  #1476·diwakergupta opened Oct 27, 2026·9 tasks done·

  ·

• exl3: opt dense qwen3_5 prompts in to 64-row tensor-core tiles (stacked on #1474)

  #1475·diwakergupta opened Oct 27, 2026·9 tasks done·

  ·

• qwen3_5 cuda: serve dense EXL3 checkpoints with the DFlash2 drafter

  #1474·diwakergupta opened Oct 27, 2026·9 tasks done·

  ·

• qwen3_5: a modelopt checkpoint refuses below sm_90 in words, not missinglanepart

  #1470·javierbits opened Oct 27, 2026Contributor·9 tasks done·

  ·
  1

  1

• exl3 cuda: the 27B's EXL3 speed series on the native Dense module (#260), stacked on #1465

  #1467·BHCC2025 opened Oct 27, 2026Contributor·8 tasks done·

  ·
  1

• Nemotron: two members a pass at 8 rows a simdgroup for wide M1 windows

  #1466·mrv777 opened Oct 27, 2026Contributor·

  ·
  3

• qwen3_5 cuda: dense EXL3 checkpoints (Qwen3.8-27B), with DFlash2 drafts

  #1465·BHCC2025 opened Oct 27, 2026Contributor·8 of 9 tasks·

  ·
  2

• flashnext: read PLE rows concurrently through readMany

  #1462·paragontasx opened Oct 27, 2026Contributor·7 of 9 tasks·

  ·
  1

• fp8 experts: prompt tiles copy weights into contiguous shared memory (-19 to -28 % at T <= 2048, bit-equal)

  #1460·jschmied opened Oct 27, 2026Contributor·

  ·
  3

• moss_image: MOSS-Image-Base's transformer on the Metal runtime (new family, denoiser)

  #1457·BobClawblaw opened Oct 27, 2026Contributor·9 tasks done·

  ·
  1

  4

• EXL3 Flash Next: draft lane guard for offloaded experts (8/8)

  #1446·BlivionIaG opened Oct 27, 2026Contributor·

  ·
  1

• EXL3 Flash Next: CUDA graphs and the device planner (7/8)

  #1445·BlivionIaG opened Oct 27, 2026Contributor·

  ·
  1

• EXL3 Flash Next: KV pages from the core pool (4/8)

  #1444·BlivionIaG opened Oct 27, 2026Contributor·

  ·
  1

• EXL3 Flash Next: deep prefill (3/8)

  #1443·BlivionIaG opened Oct 27, 2026Contributor·

  ·
  1