All pull requests

Pull requests

• cuda flashnext: --learn-conversations keeps Flash Next's states on disk, and CUDA serves --spill-gib (#840)

  #1359·jschmied opened Oct 26, 2026Contributor·

  ·

• cuda qwen3_5: --learn-conversations keeps the 27B's and 35B's states on disk (#840)

  #1358·jschmied opened Oct 26, 2026Contributor·

  ·

• cuda qwen3_5: the prompt cache keeps KV pages on the device (TF_QWEN_KV_SHARE, #1156)

  #1357·jschmied opened Oct 26, 2026Contributor·

  ·

• cuda qwen3_5: an MTP stream's ring and state are part of its pool lane's run (#1156)

  #1356·jschmied opened Oct 26, 2026Contributor·

  ·

• cuda qwen3_5: a paged prompt chunk's attention goes through most_verify rows at a time

  #1355·jschmied opened Oct 26, 2026Contributor·

  ·

• nemotron CUDA: the router top-k is moe_glue's sigmoid top-k (#548, step 2)

  #1354·jschmied opened Oct 26, 2026Contributor·

  ·

• nemotron CUDA: RMSNorm rows run on the shared norms and combine (#548, step 1)

  #1353·jschmied opened Oct 26, 2026Contributor·

  ·

• lane host: a long prompt's step owed to decoders ends on its device

  #1352·Bizuayeu opened Oct 26, 2026Contributor·9 tasks done·

  ·

• nvfp4 experts: Flash Next decode rows in equal waves, two warps a unit for one row

  #1351·jschmied opened Oct 26, 2026Contributor·

  ·

• prompt16: persistent 128x256 FP4 tile with a decoupled copy ring

  #1350·jschmied opened Oct 26, 2026Contributor·

  ·

• flashnext: shared sessions spill evicted states to disk (--spill-gib)

  #1349·gilby opened Oct 26, 2026Contributor·

  ·

• qwen CUDA: presence_penalty and frequency_penalty, each verify row over its reply and its own draft path

  #1346·BobClawblaw opened Oct 26, 2026Contributor·

  ·

• flashnext: a shared round draws a sampled stream's drafts under its keyed rule

  #1343·gilby opened Oct 26, 2026Contributor·

  ·

• flashnext: a shared verify runs the sessions' recurrences without barriers between sessions

  #1342·gilby opened Oct 26, 2026Contributor·

  ·

• lanes: a batched head's extra drafts are priced as the deepest stream's steps

  #1341·gilby opened Oct 26, 2026Contributor·

  ·
  1

• flashnext: shared rounds draft every stream in one packed MTP pass a step

  #1340·gilby opened Oct 26, 2026Contributor·

  ·

• flashnext: the kept-state copies go into the verify command buffer, fenced before the forward

  #1339·gilby opened Oct 26, 2026Contributor·

  ·

• flashnext: a shared round samples every sampled window in one dispatch

  #1338·gilby opened Oct 26, 2026Contributor·

  ·

• lanes sampling: a keyed draw that cuts nothing by rank finds the sorted draw's token in one pass

  #1337·ajensenwaud opened Oct 26, 2026Contributor·9 tasks done·

  ·
  1

• server: replies written as messages to recipients split into reasoning, answer and tool calls

  #1336·ajensenwaud opened Oct 26, 2026Contributor·9 tasks done·

  ·

• qwen3_5 cuda: a chat's next turn resumes at the previous turn's history (message starts for the planned cache)

  #1335·gilby opened Oct 26, 2026Contributor·

  ·
  1

• Sage-A1.5 (Qwen3.5 MoE 35B-A3B) on the XPU backend: decode, serve, vision and MTP

  #1331·tolewis opened Oct 26, 2026·7 of 9 tasks·

  ·
  1

• xpu: EXL3 trellis weights on Intel GPUs (loader names, kernels, Level Zero host, checks)

  #1330·tolewis opened Oct 26, 2026·8 of 9 tasks·

  ·
  1

• lanes: --decode-share sizes a filling prompt's chunks beside live replies (#840)

  #1327·jvr0x opened Oct 25, 2026Contributor·

  ·
  2

• Muse Spark Open on CUDA: EXL3, DFlash2, two GPUs

  #1295·ajensenwaud opened Oct 25, 2026Contributor·8 of 9 tasks·

  ·
  1