DeepSeek-v4-Flash

DeepSeek-v4-Flash

Pulsar's build of DeepSeek-v4-Flash, derived from deepseek-ai/DeepSeek-V4-Flash-Vision-Exp at revision 6821d6ad3681a4b137b066b76094fa82ebd0a380.

This is the unqualified member of the family: the source-precision build. Every routed expert is stored in cutlass_mxfp4, repacked from the upstream checkpoint's own FP4 bytes — nothing is quantized down, and there is no iq2_xxs_mmq_k tensor anywhere. Variants that deviate from source precision carry a suffix (-IQ2, -Mix-…); the bare name always means this one.

At ~168 GB it does not fit a single GB10. It is the artifact for a two-Spark rig; the 2-bit build (DeepSeek-v4-Flash-IQ2, ~92 GB) is the single-Spark one.

Layout

Engine-layout tensors are stored as flat U8 blobs. Each shard's __metadata__ is self-describing:

• pulsar.tensors — per tensor: layout and dims_ne (the logical shape in ggml ne[] order, which is the HF shape reversed; recorded for every tensor so a reader never has to reason about orientation).
• pulsar.experts — per routed-expert projection: n_experts, expert_bytes, layout, contiguous, and the original stacked tensor name. The expert tensors of a projection are written back-to-back with no inter-expert padding, so a kernel can address them as base + xid * expert_bytes from a single pointer.
• pulsar.kv_arch — the 53 architecture keys the engine reads. pulsar.kv (primary shard only) additionally carries the tokenizer block.
• format — pt, and pulsar.alignment — 32. Every tensor offset is aligned to 32 bytes, which the swizzled scale tiles depend on.

Sizes

The expert payloads are byte-count identical — 157,437,394,944 B in both — and a pure permutation: the upstream E2M1 nibble plane is copied verbatim and the upstream E8M0 group-32 plane is scattered into the CUTLASS SFB swizzle. The +155 MB is the dense scale plane: mxfp8_lt carries one E8M0 scale per 32 columns, where the upstream FP8 weight block is 128×128. That re-encode is lossless to within E4M3 rounding on 0.0013% of dense elements (max |Δ| 1.5e-5).

Loading

This runs on pulsar only. Stock transformers / vLLM cannot load it — they have no decoder for the fused layouts, and U8 blobs have no meaning without the metadata above. config.json's quantization_config.quant_method is pulsar-safetensors precisely so a standard loader fails loudly instead of misreading the bytes.

Verification

• 48 shards, 36,909 tensors, index closed both ways, gap-free, 32 B aligned, exact file lengths.
• 138/138 routed-expert projections cutlass_mxfp4 and contiguous.
• Zero shards contain the substring iq2.
• 45/45 spot-checked expert blobs invert to the upstream E2M1+E8M0 bytes exactly.
• Non-expert tensors byte-identical to the base container (48/48 shard digests).
• Dense values vs upstream: 83,326 of 6,304,038,912 elements differ, max |Δ| 1.5e-5.

See BUILD-RECORD.md for the full provenance and gate list.