biondizzle 16b72b9581 PERF: Eliminate double quantization for o_a_proj + NVFP4 lm_head ...

1. o_a_proj (Nvfp4GroupedLinear): Added load_nvfp4_weight() method
   that loads checkpoint NVFP4 weights directly — no more dequant→BF16→requant.
   Each group's weight is transposed from (N, K_packed) checkpoint layout
   to (K_packed, N) layout expected by the grouped GEMM.

2. lm_head: Quantize BF16 weight to NVFP4 at load time, use production
   Nvfp4Linear GEMM instead of F.linear. Runtime gsa for activation.
   Frees the 1.8GB BF16 weight after quantization.

3. Hash router (L0-2): Already optimal — tid2eid is an int32 lookup,
   no GEMM to accelerate.

2026-06-01 19:41:21 +00:00

..

__init__.py

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

attention.py

E2/E3: compressor bridge, indexer bridge, flush pipeline wiring

2026-05-30 21:16:54 +00:00

embedding.py

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

ffn.py

Layer dispatch: config, schedule, attention/FFN sub-blocks, TransformerLayer

2026-05-21 23:11:09 +00:00

grouped_linear.py

PERF: Eliminate double quantization for o_a_proj + NVFP4 lm_head

2026-06-01 19:41:21 +00:00

linear.py

CRITICAL FIX: runtime activation global scale to prevent E4M3 overflow

2026-06-01 14:21:16 +00:00

mhc.py

Add --no-thinking mode to skip thinking tokens and use second-best

2026-05-31 19:24:21 +00:00

moe.py

Fix MoE non-fused L2 runtime gsa + update test harness for extra args

2026-06-01 15:03:54 +00:00

norm.py

Fix layer construction: match existing API signatures, add RMSNorm impl

2026-05-21 23:31:58 +00:00

router.py

CRITICAL FIX: runtime activation global scale to prevent E4M3 overflow

2026-06-01 14:21:16 +00:00

shared_expert.py

CRITICAL FIX: runtime activation global scale to prevent E4M3 overflow

2026-06-01 14:21:16 +00:00