native/cuda_memory: count every file the weight map names in weightBytes

[Bizuayeu]

What this changes

weightBytes, the estimate that refuses a launch that cannot fit, now counts each file that model.safetensors.index.json's weight map names, once, so shards in a subfolder are no longer left out (#1015). Without an index, or with an index that has no weight map, it counts the folder's own *.safetensors as before. In 1.1.0 the estimate is called from native/cuda_family_memory.zig, which now passes its allocator.

• nvidia/GLM-5.3-Flash-FP4-Local keeps one shard in tensors/ (tensors/native-QB.safetensors).
• nvidia/Qwen3.8-Flash-Next-FP4-Local keeps all of its shards there.

Receipt

• Environment: TensorFold 1.1.0 (76aafcf2) plus this commit, Zig 0.17.0. Host tests on x86_64 Linux, on an LF tree from git archive. The byte counts below are the files' sizes on a GB10.
• The estimate before and after:
  • nvidia/GLM-5.3-Flash-FP4-Local@76348ddb: 176,710,148,760 → 177,463,841,024 bytes. The difference, 753,692,264, is tensors/native-QB.safetensors.
  • nvidia/GLM-5.3-Flash-NVFP4@da920bb: 204,439,103,396 both ways. All 33 of its files are in the folder.
• Exactness: n/a. Only the admission estimate changes.
• Decode speed before and after: n/a, for the same reason.
• Prompt speed before and after: n/a, for the same reason.
• Tests run:
  • zig build test: 152/152 steps, 2414/2423 tests passed, 9 skipped.
  • zig build test-kernel-contract, test-golden, fuzz-cortheon, the tools/zig unittest (flashnext_batch_serve_test.py) and a plain zig build: succeeded.
  • lean: 0 new problems (503 on main). git diff --check: clean.
  • The new test covers three cases: no index (the folder's own files, not the subfolder's); an index (each named file once, the subfolder's too); and an index without a weight map (the folder's own files).
• Not run: a CUDA launch, since the estimate is checked by the host test and the byte counts above. Metal, since the file is CUDA's.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (Only the admission estimate changes.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (No inference path changes.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (No arithmetic changes.)
• 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 index is read once at admission.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not. (Host tests on x86_64 Linux; the byte counts on a GB10; no CUDA launch, no Metal.)
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (Before it, the indexed case counts 110 bytes instead of 17.)
• 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 finding the missing 753,692,264-byte shard in the admission estimate, while the existing NVFP4 checkpoint stays at 204,439,103,396 bytes. We're taking this with fixes for 1.1.1. Distinct indexed filenames, including subfolders, are counted once. Review added refusal for missing named files, malformed indexes and allocation/stat failures, so a broken index cannot silently become a smaller flat-folder estimate. Valid no-index folders retain their old counting behavior. This changes admission accounting, not inference. The host and admission-refusal controls cover those cases; complete FP4-Local loading and rank/device checks remain under #1015.

Posted by Ash's AI Assistant Synthena

[Bizuayeu]

Rebased onto 1.1.0 (76aafcf2): the estimate's caller moved to native/cuda_family_memory.zig, so this now passes the allocator there. The receipt above is from the rebased commit.

[ashhart]

Thanks @Bizuayeu. 1.1.1 counts each indexed filename once, including subfolders, and refuses malformed indexes, missing/non-file shards, allocation/stat failures and overflow before loading. Only an absent index or weight map falls back. Your 694c62a rebase retains the original permissive reader; our maintained fail-closed equivalent already covers the required repair. Closing for the shipped accounting scope, with your credit retained. Mixed-format GLM serving remains a separate checkpoint gate.

Posted by Ash's AI Assistant Synthena