hugo commented on issue hugo/DS4Server#106 2026-10-04 11:42:37 +00:00
Commit the final MTP primary before banking warm session state

Implemented and verified in commit a8cad8810060d22ee7ccac82c210ef40d4e08785 (pushed to main).

The round-head stop/token-limit exit now keeps the emitted primary pending, matching MTPLX 2.12.2…

hugo pushed to main at hugo/DS4Server 2026-10-04 11:42:18 +00:00
a8cad88100 Commit the final Qwen MTP primary before saving state
hugo closed issue hugo/DS4Server#105 2026-10-04 11:29:45 +00:00
Reject invalid Qwen logits instead of emitting fabricated tokens
hugo commented on issue hugo/DS4Server#105 2026-10-04 11:29:45 +00:00
Reject invalid Qwen logits instead of emitting fabricated tokens

Implemented in 5ecf8ab4f31aeaef7a06b4f03fbb8d658f28c486 and pushed to main.

Qwen now rejects NaN, +infinity and all-masked logits before consuming fabricated IDs. Greedy checks validity…

hugo pushed to main at hugo/DS4Server 2026-10-04 11:29:18 +00:00
5ecf8ab4f3 Reject invalid Qwen logits and discard poisoned sessions
hugo closed issue hugo/DS4Server#104 2026-10-04 10:50:18 +00:00
Use full-vocabulary top-p law in Qwen AR pipeline
hugo commented on issue hugo/DS4Server#104 2026-10-04 10:50:18 +00:00
Use full-vocabulary top-p law in Qwen AR pipeline

Implemented and pushed in 1de78fa.

Removed the Qwen AR pipeline end to end: policy default/flag,…

hugo pushed to main at hugo/DS4Server 2026-10-04 10:49:54 +00:00
1de78fa3ef Retire the Qwen AR pipeline
hugo closed issue hugo/DS4Server#102 2026-10-04 10:27:11 +00:00
opening GLM 5.3 flash session when qwen is selected gives an error
hugo closed issue hugo/DS4Server#103 2026-10-04 06:28:57 +00:00
Analyse Qwen3.8 Flash Next changes in MTPLX and create confirmed bug issues
hugo commented on issue hugo/DS4Server#103 2026-10-04 06:28:57 +00:00
Analyse Qwen3.8 Flash Next changes in MTPLX and create confirmed bug issues

bug issues created, this analysis task is done.

hugo pushed to main at hugo/DS4Server 2026-10-03 17:10:49 +00:00
5773942c8f Name the parity oracle per model family
dcc4c516f0 Record MTPLX upgrade defect reproductions
Compare 2 commits »
hugo commented on issue hugo/DS4Server#103 2026-10-03 17:09:45 +00:00
Analyse Qwen3.8 Flash Next changes in MTPLX and create confirmed bug issues

Output list addition: #112 Port MTPLX row-map SwiGLU expert gather for Qwen prefill MoE (required route-parity port per the Qwen/MTPLX oracle assignment; bit-identical to the guarded stock chain…

hugo commented on issue hugo/DS4Server#110 2026-10-03 17:09:45 +00:00
Prevent uninitialized expert outputs for large unaligned M5 prefill chunks

Ownership update: the current MTPLX Flash-Next MoE route (row-map sort + fused SwiGLU NAX gate/up kernel, guarded stock down) is now tracked as a required Qwen route-parity port in #112. The…

hugo opened issue hugo/DS4Server#112 2026-10-03 17:09:37 +00:00
Port MTPLX row-map SwiGLU expert gather for Qwen prefill MoE
hugo commented on issue hugo/DS4Server#110 2026-10-03 15:37:33 +00:00
Prevent uninitialized expert outputs for large unaligned M5 prefill chunks

Description updated after second review (2026-10-03)

Addressed the new route observation:

  • Documented current MTPLX…
hugo commented on issue hugo/DS4Server#104 2026-10-03 15:37:32 +00:00
Use full-vocabulary top-p law in Qwen AR pipeline

Description updated after second review (2026-10-03)

Addressed the remaining gaps:

  • Corrected trigger/severity:…
hugo commented on issue hugo/DS4Server#103 2026-10-03 15:18:16 +00:00
Analyse Qwen3.8 Flash Next changes in MTPLX and create confirmed bug issues

Re-verification of revised child issues (2026-10-03)

Oracle assignment confirmed by the user:…

hugo commented on issue hugo/DS4Server#110 2026-10-03 15:18:15 +00:00
Prevent uninitialized expert outputs for large unaligned M5 prefill chunks

Re-verification of the revised description (2026-10-03) — remaining gap

The revised pad/slice rule, predicates and non-goals match b36b106 and mtplx/moe_sorted_gather.py at 9882703…

hugo commented on issue hugo/DS4Server#104 2026-10-03 15:18:15 +00:00
Use full-vocabulary top-p law in Qwen AR pipeline

Re-verification of the revised description (2026-10-03) — remaining gaps

The revised body is correct about the two root causes (top-k-renormalized top-p; PCG64 token 0 followed by a keyed…