CalamitousFelicitousness f05a3ae3e9 fix(sdnq): skip compiled dequant for fp8 weights without hardware support ...

Triton has no e4m3 conversions before sm_89, so any compiled graph
touching fp8 storage weights fails with an InductorError on Ampere.
Select the eager dequant and re-quantize paths for e4m3 weights when
the hardware cannot compile them; other dtypes keep compiled dequant.
SDNQ_ALLOW_FP8_COMPILE overrides the detection.

2026-07-12 03:44:54 +01:00

..

kernels

SDNQ handle subnormals in FP formats and cleanup Triton MM

2026-07-11 21:52:51 +03:00

layers

SDNQ handle subnormals in FP formats and cleanup Triton MM

2026-07-11 21:52:51 +03:00

packed_int

lint

2026-06-02 17:13:08 +02:00

__init__.py

Refactor SDNQ quantizer handling and add modules_to_not_use_matmul

2026-05-15 05:47:32 +03:00

common.py

fix(sdnq): skip compiled dequant for fp8 weights without hardware support

2026-07-12 03:44:54 +01:00

dequantizer.py

fix(sdnq): skip compiled dequant for fp8 weights without hardware support

2026-07-12 03:44:54 +01:00

file_loader.py

full codespell coverage

2026-06-04 12:36:10 +02:00

forward.py

SDNQ Add UINT8 MM support

2026-07-04 18:51:45 +03:00

loader.py

SDNQ handle subnormals in FP formats and cleanup Triton MM

2026-07-11 21:52:51 +03:00

packed_float.py

SDNQ handle subnormals in FP formats and cleanup Triton MM

2026-07-11 21:52:51 +03:00

quant_utils.py

remove math import

2026-07-09 14:14:50 +03:00

quantizer.py

SDNQ handle subnormals in FP formats and cleanup Triton MM

2026-07-11 21:52:51 +03:00

utils.py

Inductor breaks down with math.log

2026-07-09 14:06:02 +03:00