Rewrite single_shot_inference.py: correct weight keys, NVFP4 two-level scale, compressor+indexer connected

Browse Source

- Fixed weight key format: model.layers.{li}.self_attn.* (was layers.{li}.attn.*)
- Added NVFP4 two-level scale: weight_scale * weight_scale_2 * input_scale
- Proper CSA compressor: overlapping Ca/Cb streams, token-level softmax
- Proper HCA compressor: non-overlapping, single stream
- Indexer: NVFP4 q_b_proj + weights_proj + own compressor at index_head_dim
- Compressed KV (dim=hd) concatenated with SWA KV for attention
- Correct MoE key format: gate_proj/up_proj/down_proj
- Correct mHC key format: attn_hc.{fn,base,scale} and ffn_hc.{fn,base,scale}
- No more disconnected compressor — full E2E pipeline

...

This commit is contained in:

biondizzle

2026-05-31 21:48:59 +00:00

parent 4988e77179

commit eb08cd06d1

1 changed files with 378 additions and 744 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

1122

single_shot_inference.py

File diff suppressed because it is too large Load Diff