Pin llama.cpp cde002d: fix NaN after MTP rollback on Qwen3.5/3.8 (llama.cpp #52)

[bong-water-water-bong]

Moves the llama.cpp pin from 764a256 to cde002d, the merge of llama.cpp #52.

The gated delta-net snapshot path now forms its decay ratios in log space. Previously, exp(G_t) / exp(G_s) became 0/0 = NaN once a strongly forgetting head underflowed. This fixes two failures on HRX:

• Qwen3.8-27B with --mtp aborting with "NaN logits at vocab index 0" on some prompts (Q8_0, UD-Q5_K_XL, UD-Q6_K files);
• an AMDGPU memory fault on a repeated request without MTP.

Results, from #52:

• Q8_0 with --mtp: 0 NaN on prose and code (15.7 / 20.2 tok/s).
• Repeated request without MTP: clean.
• UD-Q4_K_XL with --mtp: speed, draft acceptance and output text unchanged.

Other changes in this PR:

• The registry is regenerated; only the pin line changes, and --check-pins passes.
• docs/hrx.md gets a note on the fix.

🤖 Generated with Claude Code

[bong-water-water-bong]

Checked:

• Submodule: 764a256..cde002d contains only llama.cpp #52 and its merge: one file, gated_delta_net_f32_wmma.loom, +28 / -12. cde002d is the current tip of 1bit/hrx-vulkan-patched (checked with ls-remote).
• Registry: the llama.cpp (hrx) line matches the new pin, and nothing else changed.
• docs/hrx.md: the description matches what was measured, which is the 0/0 decay ratio in snapshot publishing, the prompt dependence, the affected files, and the memory fault on a repeated request without MTP.

One small suggestion for the note: the speeds (15.7 / 20.2 tok/s, and the UD-Q4_K_XL comparison) were measured in balanced power mode (85 W), so it's worth saying so next to the numbers, since the other figures in this file are from performance mode. Not blocking.

Fine to merge once checks pass.

🤖 Generated with Claude Code

[github-actions]

Preparing review...

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis ✅ 52 - Fully compliant Compliant requirements: Pin llama.cpp to cde002d commit Fix NaN after MTP rollback on Qwen3.5/3.8 Update registry with new llama.cpp commit Document the fix in docs/hrx.md

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation mentions that the fix resolves NaN logits at vocab index 0 and an AMDGPU memory fault, but does not specify the exact conditions under which these issues occurred or how to reproduce them. This lack of detail may make it difficult for others to understand the scope of the problem or verify the fix. - **NaN after an MTP rollback on Qwen3.5/3.8 ([llama.cpp #52](https://github.com/1bit-MONSTER/llama.cpp/pull/52)).** With `--mtp`, the gated delta-net kernel publishes one recurrent-state snapshot per verify token so a rejected draft can roll back. It formed each snapshot's decay ratio c_t/c_s by dividing two exponentials. When a strongly forgetting head drove both to 0, that gave 0/0 = NaN, and the next verify resuming from that snapshot returned NaN logits ("NaN logits at vocab index 0"). It depends on the prompt, and showed up on Qwen3.8-27B Q8_0, UD-Q5_K_XL and UD-Q6_K files. The same bug caused an AMDGPU memory fault on a repeated request without MTP. The ratios are now formed in log space, as the main recurrence already did. Q8_0 `--mtp` now runs clean on prose and code (15.7 / 20.2 tok/s, balanced power mode, 85 W); UD-Q4_K_XL `--mtp` speed and output are unchanged (also measured in balanced mode).

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	⚠️ Did not finish.	View run

Commit 4a1c82b