[jayleaton]

What this changes

tool_choice: "required" and named tool calls compile a structural-tag grammar from the template's call envelope. DeepSeek-V4.x writes calls in DSML (<｜DSML｜invoke …>), and the grammar had no DSML style, so these requests could not be constrained on DeepSeek. This adds a deepseek style. It recognizes the native template's call envelope and emits structural tags with each offered tool's argument schema. Named-tool filtering and the parallel-call policy are kept. The grammar envelope is parsed in whole replies and in streamed prefixes.

Issue: none open upstream. Related: #1396 (stream and whole-reply agreement, including DSML whitespace), and #1665 (DeepSeek-V4.1 grammar masks on CUDA), which does not touch the tool-call grammar this PR changes. This PR does not change the stream filter that #1396 covers. Stacked on: nothing (based on main d2ad5a7f2).

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision: head e9f668512bb2 on main d2ad5a7f2; Zig 0.17.0; host checks on a Linux aarch64 DGX Spark (GB10, CUDA 13.0). Real-pack boot on two GB10s, TP=2 over RoCE, checkpoint jayleaton/DeepSeek-V4.1-Flash-1017-EXL3-2.8bpw revision 859396591c34, TF_EXPERIMENTAL=deepseek.
• Exactness: n/a; other styles are unchanged, and DeepSeek required calls go through the same grammar path as the other styles.
• Decode speed before and after: n/a.
• Prompt speed before and after: n/a.
• Host tests (tool_grammar_deepseek_test.zig): the actual template, schema alternatives, named calls, typed arguments, and every streaming byte boundary of the envelope.
• Tests run, with pass, fail and skip counts:

  zig fmt --check build.zig zig                      exit 0
  git diff --check d2ad5a7f2..HEAD                    exit 0
  python3 tools/zig/lean_check.py                     base 424 problems, head 424, new 0 (lean_new.py)
  zig build test-cuda-host --summary all              exit 0, 273/273 steps, 4081/4090 tests passed, 9 skipped (with test-server-cpu)
  zig build test-server-deepseek-chat-host test-tool-visibility-http-host test-golden  exit 0, 11/11 steps, 45/45 tests passed
  

• Not run, and why: a device run with xgrammar on the real pack.
• Overlap: atc overlap reports one-line additive conflicts with minimax_m3: a CUDA backend for MiniMax-M3.1-Flash at TP over GB10s, from the release's own formats #1388, minimax_m3: serve MiniMax-M3.1-Flash (428B-A23B MoE, images) natively on Metal #1235 and minimax_m3 (CUDA): draft several streams' DSpark blocks in one pass #1426 (MiniMax-M3), which each add another Style case and leads entry in tool_grammar.zig and reply_text.zig. Whichever lands second keeps both entries.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Not measured for this change alone.)
• No precision traded for speed. If the bits change, the description says which sums change and why.
• 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 kernel changes.)
• Prompt processing is no slower than the last release. (Server grammar only.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No reverted-fix control was run for this PR.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[jayleaton]

Rebased onto main d2ad5a7f2 to resolve two conflicts with main's MiniMax-M3 tool calls. In reply_text.zig, main's Openers table and MiniMax block are kept, and the DSML function_calls opener is added after the existing openers, so the indices MiniMax and LFM2 use (calls[2]) are unchanged. In tool_grammar.zig, Style now has both minimax and deepseek; call() already handled both. The host gates were re-run on the new head and pass; the description has the new numbers.

[ashhart]

Thank you @jayleaton. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @jayleaton! Required and named tool calls held to DeepSeek's DSML markup are in 1.2.1: https://github.com/ashhart/TensorFold/releases/tag/v1.2.1