feat(deepseek_v4): PR1 skeleton — end-to-end inference with triton MoE#650

Merged

valarLip merged 48 commits intomainfrom

feat/deepseek-v4-pr1-skeleton

May 6, 2026

Commits on Apr 25, 2026

refactor(deepseek_v4): switch forward to ATOM 2D flat-token convention
valarLip
committed
feat(deepseek_v4): swap MoE to FusedMoE for 384-expert TP/EP loading
valarLip
committed
fix(deepseek_v4): V4QuantConfig now matches FusedMoE's bare 'experts' prefix
valarLip
committed
fix(deepseek_v4): correct FusedMoE expert weight + scale + bias dispatch
valarLip
committed
feat(deepseek_v4): wire hash routing for first 3 layers via custom_routing_function
valarLip
committed
feat(deepseek_v4): full Block.forward (attn + FusedMoE) end-to-end on real ckpt
valarLip
committed
feat(deepseek_v4): standard ATOM loader (load_model) now handles V4 ckpts
valarLip
committed
fix(deepseek_v4): wo_a FP8 dequant via process_weights_after_loading hook
valarLip
committed
feat(deepseek_v4): end-to-end inference with triton MoE and swiglu_limit
valarLip
committed

Commits on Apr 29, 2026

Commits on May 2, 2026

Commits on May 4, 2026

Commits on May 5, 2026

Commits on May 6, 2026