Pin llama.cpp f5b7f4a: release pin (gpt-oss, ternary, MLA, deterministic FA, attention sinks)

[bong-water-water-bong]

Release pin for Sunday 2026-10-04: llama.cpp fork 2bd7f58 -> f5b7f4a. Three commits, one per step:

• 2bd7f58:
  • 1bit comfy: run the checked binary; fetch the submodule when missing #66/Pin Lemonade: third_party/lemonade at the fork's 7650b4f #67/NPU route: reasoning_content, as llama-server returns it #73 gpt-oss on HRX: MUL_MAT_ID for sizes that are a multiple of 32, ADD_ID and SWIGLU_OAI, and a placement guard.
  • 1bit serve: forward llama-server's other routes (Lemonade checklist #3) #69 TQ1_0 / TQ2_0.
  • 1bit serve --embedding / --reranking (Lemonade checklist #4) #70 MLA prompts past 512 tokens.
  • Bump Lemonade to d7c644a9; embedding checklist status #71 Hadamard qwen3next.
  • NPU: the build's own lax kernels are the default; Qwen3.6-35B-A3B through Lemonade #72 deterministic flash attention with masked keys.
• 4485916: docs/lemonade.md: the embedding checklist (measured) #68 attention sinks on HRX.
• f5b7f4a: README, NOTICE: acknowledge FastFlowLM #74 PrismML tile bytes (PQ2_0 small-model decode).

Each commit updates docs/hrx.md "Our patches" and the registry pin.

Release format matrix through 1bit serve --device hrx (balanced mode). All 18 rows load with no CPU fallback, and repeated greedy answers are identical. Highlights (decode tok/s):

format	model	tok/s
Q4_K_M	Qwen3-0.6B	317
Q1_0	Bonsai-1.7B	194
Q1_0	Bonsai-8B	56
TQ1_0	Ternary-Bonsai-1.7B	114
TQ2_0	Ternary-Bonsai-1.7B	159
PQ2_0	Ternary-Bonsai-1.7B	206 (CPU-identical answer, at f5b7f4a)
PQ2_0	Ternary-Bonsai-2-27B	19.2
PTQ1_0	Ternary-Bonsai-2-27B	16.0
UD-Q4_K_XL	Qwen3.8-27B	11.9
MLA Q4_K_M	GLM-4.7-Flash, 761-token prompt	21
MXFP4	gpt-oss-20b	39 (at f5b7f4a)

• gpt-oss-20b vs CPU: perplexity 463.0 vs 459.0; mean KLD 0.027 (max 1.99); same top token 87.4%. Our claim is "runs on HRX, KLD 0.027 vs CPU", not bit-close.
• Too few bits for a 0.6B: Qwen3-0.6B IQ1_S / IQ2_XS / Q2_K give the same poor text on CPU. They load and run, and the quality is the model's.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit d91683b