hanzo-dev 0c4edccb40 feat(v4): DeepSeek-V4-Flash DSpark training pipeline against a quantized serving target ...

convert_engine_dump.py: hanzo-engine hidden-state dumps -> TARGET_CACHE_VERSION=2 (reuses DeepSpec's own writers; bit-exact round-trip). dspark_v4flash_quant.py: V4-Flash draft config (hidden 4096, vocab 129280, 43 target layers, layer_ids [1,11,21,31,41], block 7, markov vanilla 256, confidence head, mask_token 129279). v4_pipeline/: synthetic dump generator + training smoke (single-GPU, reduced dims). PROVEN on GB10: 300 real samples dumped at 111.5 tok/s prefill -> converted -> 50 steps on real quantized-V4 hiddens, loss 3.03->2.6 decreasing; full config 5L/64a/12288 passes solo.

2026-07-02 21:20:04 -07:00

..

data

feat(v4): DeepSeek-V4-Flash DSpark training pipeline against a quantized serving target

2026-07-02 21:20:04 -07:00

eval

Align DSpark/DFlash exp_name with block_size=7 (block7)

2026-06-29 21:23:50 -07:00

train

first init

2026-06-27 12:56:53 +08:00