[bong-water-water-bong]

Fixes #95. Supersedes #101 and #103: their commits are in this pin, plus the fix that #103 needed.

Pin. third_party/llama.cpp moves c075cc1 -> 96f6b89 (1bit-MONSTER/llama.cpp#4, merged into 1bit/hrx-vulkan-patched):

• d7ff3af (from hrx: fix non-FA attention path for MoE models (500 "Compute error") (#95) #101): the non-FA MoE graph compile failures and the sigmoid-router claim.
• 07a2ee4 (from hrx: claim flash attention for MLA and full-context MoE shapes (#95) #103): HRX0 claims flash attention for MLA (head 576) and at full context (262144).
• ff07f1c: MUL_MAT_ID reads route ids at their row stride. This was the GLM batched wrong answer. llama.cpp's route ids are an [n_used, T] view of the [n_expert, T] argsort, and HRX read them as packed.
• d925875: the FA value_stride range covers a whole KV row. As posted, hrx: claim flash attention for MLA and full-context MoE shapes (#95) #103 broke Qwen3-0.6B, with value_stride 1024 over the range max of 576.
• 19465d2: the GGML_HRX_DISABLE_DISPATCH / GGML_HRX_LOG_DISPATCH debug switches.

Engine. Removes #97's 32768 context default for --device hrx: HRX0 now keeps flash attention at a model's full context.

Verified on Strix Halo, fresh build of this branch (-DONEBIT_HRX=ON):

• tools/registry_check.py --only hrx: 7 of 7 pass at full context. That covers Qwen3-0.6B, Qwen2.5-7B, Qwen3-Coder-30B-A3B, Qwen3.6-35B-A3B Q8_0, GLM-4.7-Flash, MiniCPM4-8B and MiniCPM5-1B.
• serve_e2e ... vulkan --prefill-device hrx passes.
• test-backend-ops -b HRX0: 790/790.

Docs: registry.md, PORTING and serve.md are updated, and registry/checked.json has the new HRX rows.

🤖 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 855a543

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 95 - Partially compliant Compliant requirements: Fix HTTP 500 "Compute error." when running Qwen3-Coder-30B-A3B-Instruct Q4_K_M and GLM-4.7-Flash on HRX Ensure MoE models can run at full context (262144 tokens) on HRX Resolve issues with non-FA attention path for MoE models Fix flash attention claims for MLA and full-context MoE shapes Non-compliant requirements: None Requires further human verification: Verification that the fixes work as expected on actual hardware (Strix Halo) Confirmation that the batched MoE routing bug is fully resolved (GLM-4.7-Flash answer correctness)

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Context Size Default Change The default context size for --device hrx has been changed from 32768 to allow full context. This change removes the limitation that caused flash attention to be disabled at full context for MoE models, which was the root cause of the 500 "Compute error." This is a functional change that impacts how models are served on HRX. if (o.ctx_size > 0) { argv.push_back("-c"); argv.push_back(std::to_string(o.ctx_size)); } Registry Update The registry documentation has been updated to reflect that GLM-4.7-Flash and Qwen3-Coder-30B-A3B now pass on HRX after the llama.cpp pin. This is a documentation update that correctly reflects the state of the system after the fix. GLM-4.7-Flash and Qwen3-Coder-30B-A3B failed on HRX until the llama.cpp pin `96f6b89` (#95). The pin fixes four things: - HRX0 now claims flash attention for MLA heads (576) and at full context (262144). - The path without flash attention now compiles. - HRX0 no longer claims a sigmoid MoE router. - HRX0's expert matmul (MUL_MAT_ID) now reads each token's expert ids at their real row stride. It used to read them packed, so every token of a batched prompt after the first went to the wrong experts. Qwen MoE models were immune, because their router dispatch supplies the stride.