[Bizuayeu]

What this changes

Adds core/nvfp4_csf32.zig, a host decoder and encoder for csf_32x2_v1, the lossless encoding of the routed experts' NVFP4 block scales in NVIDIA's GLM-5.3-Flash-FP4-Local (#1015). Decoding gives the e4m3 [n, k/16] bytes that core/nvfp4.zig reads today; nothing calls it yet.

• The decoder takes the five stored parts and allocates nothing. One pass checks them against the format and refuses malformed parts: wrong lengths, a base above 240, page prefixes that do not start at 0 or that decrease, a page origin that does not follow the records before it, records out of their block or out of order, and records left over.
• The encoder chooses each row's base as the format's reference does: the 16-wide window that covers most of the row's bytes, the lowest on a tie.
• The format is docs/csf-scale-blocks.md of local-inference-lab/flashinfer (Apache-2.0). The code is written for TensorFold from that document. The test's golden digests were computed with its reference encoder, scripts/encode_csf.py (Apache-2.0). None of its code is in this change.
• The file sits beside core: decode lossless NVFP4 CSF scale planes #904's slab codec, core/nvfp4_csf.zig, which reads the slab encoding (csf, Qwen3.8 FP4-Local). This file reads the block one, csf_32x2_v1, and takes its name from the checkpoint's weight_scale.csf32_* tensors. The two share no code.

Receipt

• Environment: TensorFold 9ccf4727 plus this commit, Zig 0.17.0. Host tests on x86_64 Linux, on an LF tree from git archive. The checkpoint check ran on a GB10 (aarch64) against nvidia/GLM-5.3-Flash-FP4-Local@76348ddb.
• Exactness: n/a, since no served path calls this. On the checkpoint, all 36,288 routed-expert scale matrices were decoded and encoded again: 24,192 of 2048 × 256 and 12,096 of 4096 × 128, with 19,025,362,944 scales and 53,886,672 exceptions. Every matrix gave back its five stored parts byte for byte. The check is a ReleaseSafe tool built on this file, and it also checks that each weight is u8 [R, K/2] with K = 16C.
• Decode speed before and after: n/a, host code that no served path calls.
• Prompt speed before and after: n/a, for the same reason.
• Tests run:
  • zig build test: 84/84 steps, 1383/1390 tests passed, 7 skipped.
  • zig build test-kernel-contract, test-golden, fuzz-cortheon and a plain zig build: succeeded.
  • lean: 0 new problems (517 on main). git diff --check: clean.
  • nvfp4_csf32.zig adds 7 tests:
    • a block built by hand;
    • the format's tile order;
    • the base choice, with ties and the 240 cap;
    • the reference encoder's bytes for three boundary shapes (32 × 1, 160 × 3, 160 × 67);
    • round trips of seven shapes, up to 2048 × 256 and 4096 × 128;
    • malformed parts refused;
    • a page's origin.
  • A copy with the tile order changed to row-major still passed the round trips, but failed the order and golden tests.
• Not run: GPU checks, since there is no kernel and no served path. Metal, since there is no Mac here.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (Host code only; no serving code changes.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (No served path changes; nothing calls the decoder yet.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (Decoding is lossless: it returns the stored e4m3 bytes exactly.)
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. (No GPU kernel.)
• Prompt processing is no slower than the last release. (No served path changes.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not. (Host tests on x86_64 Linux and the checkpoint check on a GB10; no Metal.)
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (They need only std; a changed tile order fails the order and golden tests.)
• Comments and docstrings are one line. No measurements or history in the source. (lean: 0 new problems.)
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[ashhart]

Thanks @Bizuayeu for the block-CSF codec and the reported byte-for-byte round trips of all 36,288 scale matrices. We're taking this with a geometry/exception-count overflow guard for 1.1.1. The independent order/digest tests matter: a matching encoder and decoder can both have the wrong layout and still round-trip. Review keeps the local-inference-lab format/reference attribution and checks allocation-failure cleanup. The full checkpoint result remains your reported receipt; we have not independently repeated it. This is a host foundation for #1015, with the mixed-format loader and its device/model gate still separate. No further codec changes are requested beyond the guarded integration.

Posted by Ash's AI Assistant Synthena

[ashhart]

Thanks @Bizuayeu for #1021 and your reported round trip of all 36,288 scale matrices. 1.1.1 includes the block-CSF host codec with the geometry, code-length and exception-count overflow guards, independent layout/order checks and allocation cleanup. The local-inference-lab format and reference attribution stays with it. This is a host foundation; mixed-format family loading and #1015's checkpoint/device/model checks remain separate. The full checkpoint round trip has not been independently repeated here.

Posted by Ash's AI Assistant Synthena