WIP: DeepSeek-V4-Flash architecture port (BF16 reference, accuracy-first MVP) - #14

Draft

danielhanchen wants to merge 10 commits into

deepseek-v4-flash-port

Commits on Apr 29, 2026

deepseek-v4: register architecture enum, KV keys, and tensor schema
shimmyshimmer
committed
deepseek-v4: hparams loading, minimal tensor loader, stub graph
shimmyshimmer
committed
deepseek-v4: gguf-py constants, KVs, tensor names, writer helpers
shimmyshimmer
committed
deepseek-v4: converter (DeepseekV4Model) for FP8/FP4 → BF16 dequant
shimmyshimmer
committed
deepseek-v4: chat template port (encoding_dsv4 chat-mode subset)
shimmyshimmer
committed
deepseek-v4: full per-layer tensor declaration in load_tensors
shimmyshimmer
committed
deepseek-v4 converter: only buffer FP8-packed weights for dequant
shimmyshimmer
committed
deepseek-v4 converter: emit F32 (writer downcasts) and tid2eid as I32
shimmyshimmer
committed
deepseek-v4 graph: SQRT_SOFTPLUS gating + LLM_FFN_SWIGLU_CLAMPED_PRE op type
shimmyshimmer
committed
deepseek-v4 converter: F8_E8M0 dtype + I32 numpy + tid2eid materialize
shimmyshimmer
committed