v0.6.7 — Qwen3.6 answers concurrent requests correctly

Qwen3.6 answers concurrent requests correctly, and 3.5-4x faster.

Since at least 0.5.14, any 2 or more concurrent Qwen3.6-35B-A3B requests decoded to repeated garbage tokens, for example "The capital petró fác dó fác fác …". Past 8 concurrent requests the output was correct only because the batched step was declined: requests decoded one forward each, 16 or 32 of them slower in total than a single stream. If you serve Qwen3.6 with concurrent requests, upgrade. Qwen3.8, Bonsai-2 and Muse Glimmer were not affected.

Measured

RTX 5090, qwen3_gguf_cb_bench 256/256, Qwen3.6-35B-A3B UD-Q4_K_M:

concurrent requests	0.6.6	0.6.7
4	garbage	1,055 tok/s
16	457 tok/s (one forward each)	1,621 tok/s
32	461 tok/s (one forward each)	1,834 tok/s

What changed

• Qwen3.6's packed decode step (#1262). Three faults:

  • the full-attention output gate was dropped;
  • the shared expert was not computed past 8 rows;
  • the bf16 row GEMVs (the MoE router among them) refused more than 8 rows.

  The new packed_decode_check compares packed steps with one forward per row, teacher-forced: 99-100% argmax agreement at 1-32 rows, the same as a second one-forward-per-row run (0% before).

• Bonsai-2's mixed prefill + decode steps run its FP4 FFN (#1261). A 1,024-row step takes ~86 ms instead of ~150. Continuous-batching bench at c16 / c32: 1,437 / 1,870 → 1,550 / 2,074 tok/s.

• SPARKINFER_MIXED_ROW_TOKENS (#1261): opt-in handling of request bursts for mixed steps. The CHANGELOG has the trade-off.

Full notes: CHANGELOG.md.

Container: ghcr.io/gittensor-ai-lab/sparkinfer-qwen38:0.6.7 (attested).

What's Changed

• perf(runtime): Bonsai-2's mixed steps run its FP4 FFN (cb c32 1,870 -> 2,074 tok/s); mixing a burst becomes opt-in by @skyrocket2026 in #1261
• fix(runtime): Qwen3.6 packed decode applies the attention gate and serves 9+ rows (garbage at 2+ concurrent requests; c32 461 -> 1,834 tok/s) by @skyrocket2026 in #1262
• docs: 0.6.7 — Qwen3.6 answers concurrent requests correctly (c32 461 -> 1,834 tok/s) by @skyrocket2026 in #1263

Full Changelog: v0.6.6...v0.6.7