Repository navigation

Windows ARM64 and NVIDIA RTX Spark N1X (sm_121) native CUDA support - #1010

Open

mweinbach wants to merge 30 commits into

ashhart:mainfrom

mweinbach:rtx-spark-support

Open

Windows ARM64 and NVIDIA RTX Spark N1X (sm_121) native CUDA support#1010
mweinbach wants to merge 30 commits into
ashhart:mainfrom
mweinbach:rtx-spark-support

Commits on Oct 17, 2026
- Windows: a platform layer and host builds that pass
  
  BobClawblaw
  authored and
  mweinbach
  committed
  5725ce4
- Windows: CUDA kernels, the native server and the GPU suites on sm_89
  
  BobClawblaw
  authored and
  mweinbach
  committed
  1d39fff
- Windows: pull into the hub cache, HOME, and the whole test step
  
  BobClawblaw
  authored and
  mweinbach
  committed
  ee8784c
- Windows: build notes, strip before install, formatting
  
  BobClawblaw
  authored and
  mweinbach
  committed
  f05bc1c
- Default port 8888
  
  BobClawblaw
  authored and
  mweinbach
  committed
  067e15a
- qwen3_5: serve Qwen3.5-2B on CUDA
  
  BobClawblaw
  authored and
  mweinbach
  committed
  ec24dea
- qwen3_5: serve Qwen3.8-27B on CUDA, one arena for a model's device memory
  
  BobClawblaw
  authored and
  mweinbach
  committed
  1a87bf1
- qwen3_5: replay each decode shape as a CUDA graph
  
  BobClawblaw
  authored and
  mweinbach
  committed
  a387bfe
- qwen3_5: DFlash2 drafts for Qwen3.8-27B on CUDA
  
  BobClawblaw
  authored and
  mweinbach
  committed
  7d0dd65
- qwen3_5: tree drafts on CUDA, the kept path replayed from a log
  
  BobClawblaw
  authored and
  mweinbach
  committed
  72a58ab
- docs: the Qwen hybrids on CUDA in the Windows page and the README
  
  BobClawblaw
  authored and
  mweinbach
  committed
  cf73116
- qwen3_5, platform: comments on one line
  
  BobClawblaw
  authored and
  mweinbach
  committed
  a5748b7
- Windows: checkpoint reads skip the page cache
  
  BobClawblaw
  authored and
  mweinbach
  committed
  7ca2b8d
- Windows: the signal handler type allows address 1, for aarch64
  
  BobClawblaw
  authored and
  mweinbach
  committed
  ff2910f
- docs: Windows as a target in the README and the Windows page
  
  BobClawblaw
  authored and
  mweinbach
  committed
  c083b3f
- CI: the host unit tests on Windows
  
  BobClawblaw
  authored and
  mweinbach
  committed
  3401fe4
- qwen3_5: the drafter reads TF_TIMING through the platform layer
  
  BobClawblaw
  authored and
  mweinbach
  committed
  1a07f7c
- qwen3_5: the drafter takes a prompt longer than its tap ring
  
  BobClawblaw
  authored and
  mweinbach
  committed
  0f0b711
- qwen3_5 cuda: prompt chunks past 32 rows see the right keys, and tree attention cuts chunks at logical positions
  
  BobClawblaw
  authored and
  mweinbach
  committed
  3590e8f
- Windows: the server goldens and the Cortheon fuzz step on Windows hosts
  
  BobClawblaw
  authored and
  mweinbach
  committed
  4947e2c
- cuda: cuBLASLt opens by its Windows DLL name
  
  BobClawblaw
  authored and
  mweinbach
  committed
  f76c338
- Windows: writes on an O_APPEND descriptor land at the end whole
  
  BobClawblaw
  authored and
  mweinbach
  committed
  aa81045
- golden check: a request-log mismatch names the first differing line
  
  BobClawblaw
  authored and
  mweinbach
  committed
  028e049
- golden check: the request log compares with LF endings on either side
  
  BobClawblaw
  authored and
  mweinbach
  committed
  5ae10c5
- qwen3_5: attention chunks follow each row's own keys, not the forward's widest stream
  
  BobClawblaw
  authored and
  mweinbach
  committed
  2be81c6
- qwen3_5: the DFlash2 config's divisors must be positive, and staging fits its selector rank
  
  BobClawblaw
  authored and
  mweinbach
  committed
  0f9055c
- Windows: 1.0.12's new code through the platform layer
  
  BobClawblaw
  authored and
  mweinbach
  committed
  c858400
- docs: 1.0.12's new pages at the default port, 8888
  
  BobClawblaw
  authored and
  mweinbach
  committed
  b26feda
- lanes recorder: a stream's new file is never made through a link there on Windows, where an exclusive create follows a dangling one
  
  BobClawblaw
  authored and
  mweinbach
  committed
  953bc43
- Windows ARM64 and NVIDIA RTX Spark N1X (sm_121) native CUDA support
  mweinbach
  committed
  60ba696

Commits

Commits on Oct 17, 2026