biondizzle 85e1cd3b69 Fix torch.compile crash: @torch.compiler.disable on all CuTeDSL run() ...

CuTeDSL internals (Path.cwd, threading, JIT) are incompatible with
torch.dynamo tracing. Marking run() as compiler-disabled makes the
runners opaque to torch.compile — they execute eagerly while the
rest of the model gets compiled.

2026-05-18 21:07:35 +00:00

..

patches

Wire CuTeDSL kernels into vLLM: replace all BF16 dequant with native NVFP4

2026-05-18 20:27:42 +00:00

cutedsl_quant_method.py

Fix: add abstract create_weights to CuTeDSLNvfp4LinearMethod

2026-05-18 20:40:48 +00:00

nvfp4_cutedsl.py

Fix torch.compile crash: @torch.compiler.disable on all CuTeDSL run()

2026-05-18 21:07:35 +00:00