[jschmied]

What this changes

The next step of #2460 item 3 (one sampler in core/lanes), after #2606 moved Qwen 3.5's general host row draw into
core/lanes/sampling_row.zig and #2507 (mrv777) gave Qwen's bounded top-k draw its 32-logit block skip. That bounded
draw now lives in core too:

• families/qwen3_5/draw_topk.zig moves to core/lanes/sampling_row_topk.zig (git mv, same code). The only edits
  are names: lanes.grammar/lanes.sampling become the sibling grammar.zig/sampling.zig imports, the errors
  become core's NonfiniteLogits/NoAllowedToken, the parameter sampling is settings (it would shadow the
  import), and the last line is split in two to stay under 120 columns. Normalized diff in the receipt.
• lanes.sampling_row.draw tries it first, in the same order qwen3_5/draw.zig did: a sampled top_k of 1..128
  draws from the bounded, heap-free path; greedy, top_k 0 and top_k > 128 fall through to the general path, as before.
• qwen3_5/draw.zig draw is core's draw with the errors mapped back to NonfiniteQwenLogits and
  NoAllowedQwenToken (as since lanes: the host draw over a bf16 row moves into core, Qwen 3.5 first (#2460 item 3, step 1) #2606), so every Qwen message stays the same.
• The three selection tests move with the code (zig/build/qwen35_cache_host.zig drops their names from its
  filter list: they run in core's lane tests now). A new core test pins the wiring: top_k 1, 20 and 128 through
  sampling_row.draw allocate nothing and pick the full-row choose token.

Core vs family-only:

part	where
general row draw (finite check, mask, greedy, top-k candidates, choose)	core: core/lanes/sampling_row.zig (unchanged)
bounded top-k draw (bf16 keys, 32-word block skip, stack scratch, chooseOrdered)	core: core/lanes/sampling_row_topk.zig (moved)
Qwen error names, mask_words from the Qwen vocab, Qwen full-vocabulary draw tests	family: qwen3_5/draw.zig

Not moved: HIP (hip_serve/sample.zig) and XPU (native/xpu*.zig) keep their own top-k-then-choose code. Neither is
a byte copy of this file (no bf16 key order, no block skip: f32 rows and a top_k + margin cut), so they are a later
step, not this PR.

Who else reaches the bounded path: on main only Qwen 3.5 calls sampling_row.draw. #2679 (open, step 2) moves
LFM3's draw and Flash Next's host reference onto it; once both land, their sampled top_k 1..128 rows take the same
bounded path. Its tests show the same token as the general path, so their tokens do not change.

• Issue: Refs RFC: families compose core blocks: a conformance kit first, then sampler, config, drafter and format loaders #2460 (item 3).
• Stacked on: main (223da76, which has qwen draw: a sampled row's top-k skips 32-logit blocks that cannot beat the k-th candidate #2507's 6a62797). None of the four files changed on main between 8a7838a, where this was first made, and 223da76.
• Overlap (checked by hand with gh on all 367 open PRs, then git merge-tree against this head): three open PRs touch
  these files.
  • Host draw: greedy and masked rows scan 64-logit blocks and skip mask words that allow nothing #2846 (host draw scans 64-logit blocks) changes sampling_row.zig and qwen3_5/draw_topk.zig, the file this moves,
    so the two conflict. Since this PR only moves code, Host draw: greedy and masked rows scan 64-logit blocks and skip mask words that allow nothing #2846 should land first; I then redo the move on top of it,
    so its block skip moves along unchanged, and rerun the checks below.
  • qwen3_5 metal: prefill runs a chunk a call, so a new prompt no longer stalls other streams' rounds (#3021) #3090 conflicts in zig/build/qwen35_cache_host.zig only (the test filter list).
  • logit_bias: token biases and -100 bans on each row's draw for Qwen CUDA, Qwen3.5-2B Metal and Nemotron Parse HIP #3146 (logit_bias) changes qwen3_5/draw.zig and merges with this head without conflict.
  • lanes: LFM3's row draw and Flash Next's host reference draw use core's sampling_row (#2460 item 3, step 2) #2679 (step 2) shares no file with this PR and merges without conflict.

Checks on this head

• Head commit: b0361bc
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 341 on the base (223da76), 341 on
  this head, 0 new.
• x86 (8 cores, no GPU, Zig 0.17.0) on b0361bc: zig build -Dfixed-cpu test test-golden test-kernel-contract test-server-security test-pull-security fuzz-cortheon check-release-paths: 728/728 steps, 9179 of 9246 tests passed,
  67 skipped. zig build check-native-cuda-host compile-cuda-tests: 16/16 steps. python3 tools/pr_check.py: 0 errors.
  zig fmt --check build.zig zig and git diff --check: clean.
• GB10 (aarch64), b0361bc: zig fmt --check, git diff --check and the lean checker clean; zig build native -Dgpu=spark for head and base.
• Planted controls, each on its own in a scratch copy, run against core's lane tests and test-qwen-draw-host: 13 of 13 turn red (a greedy tie taken >=, the top-k tie order, the mask ignored in either path, the position shifted in either path, three ways of breaking the 32-word block-skip bound, the block non-finite check removed, the fast path unwired, and the two Qwen error names swapped). The clean copy is green.
• CI: runs on this PR.
• Not run, and why: macOS (the moved code has no Metal dependency, and test-qwen35-metal-host runs on Linux).

Receipt

• Served binary: zig build native on the GB10 for head and base in the same path. Every embedded fatbin is identical (114 of 114), so no device code changed; the host binary differs (md5 head 277f3218, base 369e7eac).
• Served A/B, base against head, Qwen3.8-27B NVFP4 with the DFlash2 drafter, --parallel 8 --context 16384 --no-thinking, page cache dropped before each server: 11 sampling cases (greedy; top_k 1, 20 and 128 on the bounded path; top_k 0, 129 and 200 and a hot top_k 0 on the general path; JSON-schema output greedy, top_k 0 and top_k 20, where every masked row is a host draw) x 4 prompts x drafted and plain x one at a time and four at once, 48 tokens each: 176 of 176 token arrays identical, 0 differ, 0 errors.
• New switch? None.
• Environment: TensorFold b0361bc on main 223da76; Zig 0.17.0; CUDA 13 nvcc, sm_121; one NVIDIA GB10 (128 GB).
• Exactness: the same code by construction (renamed only); host tests compare each draw with lanes.sampling.choose
  over the full row, masked and unmasked, every finite bf16 word, and ties.
• Decode and prompt speed before and after: n/a. No device code changed (fatbins identical); the host draw runs the same code, with one more call.

Proposed CHANGELOG line (Unreleased, for the maintainer to place):
- Core's host row draw (lanes.sampling_row) takes Qwen 3.5's bounded top-k path; the code moved to core/lanes/sampling_row_topk.zig (#2460 item 3). Thanks to @jschmied and @mrv777.

[ashhart]

Thanks @jschmied, this is a clean next step for #2460 item 3. The bounded top-k draw moves into core/lanes as the same code with only names changed. sampling_row.draw tries it first in the order Qwen's draw already did, and Qwen's error messages are unchanged. The new test pins the heap-free path against the full-row choose token at top_k 1, 20 and 128.

It's queued to land. It's safe because on today's main only Qwen 3.5 calls sampling_row.draw, and it already took the bounded path first, so no token can change.

Optional: the three moved selection tests now run only in the full zig build test, where test-qwen35-metal-host used to run them by name. A small filtered host step for them would keep them in a fast step.

• Synthena (Ash's Agent)

[jschmied]

Thanks. So the queued head doesn't change, I'll do the filtered host step for the three moved selection tests as a small follow-up PR after this lands, rather than pushing it here.