hrx: fix non-FA attention path for MoE models (500 "Compute error") (#95)

[bong-water-water-bong]

Fixes #95.

Problem

tests/serve_e2e.sh build/1bit <model> hrx returned HTTP 500 {"error":{"code":500,"message":"Compute error."}} for MoE models whose full context exceeds HRX0's flash-attention KV cap (32768 tokens). At 262144 tokens HRX0 declines flash attention, llama.cpp disables FA for the whole model, and the non-FA attention path fails to compile on HRX0:

Model	GGUF arch	Before	After
Qwen3-Coder-30B-A3B-Instruct Q4_K_M	qwen3moe	500	200
GLM-4.7-Flash Q4_K_M	deepseek2	500	200
Qwen3.6-35B-A3B Q8_0	qwen35moe	pass	pass

Root cause & fix (in the llama.cpp submodule, 1bit-MONSTER/llama.cpp@d7ff3afb9)

Three bugs in the HRX backend's non-FA path:

1. Zero-byte alias validation — tensor_storage_relative_offset rejected a zero-byte (empty) source tensor whenever a degenerate view into it had a non-zero offset (relative_offset > ggml_nbytes(source) → 8192 > 0). Empty per-expert FFN views then failed graph-cache validation with metadata does not match. Fixed by accepting a zero-byte source when the view is also zero-byte.

2. Invalid in-place alias — the fused attention-output and routed-FFN dispatches requested an in-place buffer alias (residual_input → output) whose target is the layer output (l_out), an external/persistent buffer, so alias_storage failed with value alias target is not transient. The dispatches now only request the in-place alias when both source and target are transient, otherwise they fall back to non-fused ops.

3. Sigmoid-router SUM_ROWS — GLM-4.7-Flash's sigmoid MoE router left an unsupported SUM_ROWS on HRX0. Narrowed the SUM_ROWS/CLAMP claim to the top-k router chain (SOFT_MAX → ARGSORT → GET_ROWS), so a sigmoid router leaves those nodes to another backend.

Verification

• tests/serve_e2e.sh build/1bit <model> hrx passes (answers "Paris") on Qwen3-Coder-30B-A3B-Instruct Q4_K_M and Qwen3.6-35B-A3B Q8_0 (no regression).
• A ~2048-token prompt returns HTTP 200 (previously 500) on Qwen3-Coder-30B-A3B and GLM-4.7-Flash.

Notes / remaining gaps

• GLM-4.7-Flash still fails the answers Paris check: the 500 is gone (HTTP 200), but the reply is wrong. This is a separate correctness bug in a fused batched dispatch (correct at small ubatch, wrong at larger ubatch), being investigated in parallel — it is not the 500 "Compute error." this PR fixes.
• The explicit -fa off flag still fails on the non-FA attention postprocess (qwen attention postprocess matcher rejected node): the non-FA attention input reshape has no standalone dispatch. The default serve (FA declined at full context) exercises the non-FA path and works.

[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 e4e2a1d

[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" for MoE models on HRX backend Apply fix in the llama.cpp submodule Non-compliant requirements: Validate that non-MoE models still work (not directly tested in diff) Ensure no regression in performance or correctness (not directly tested in diff) Requires further human verification: Manual testing of the fixed models with tests/serve_e2e.sh build/1bit <model> hrx Verification that non-MoE models still work correctly

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject commit update The PR updates the llama.cpp submodule to a new commit (d7ff3afb90a5bb35bf2577fffc7d567c7ef064ed) which contains the fixes for the HRX non-FA attention path. While this is a necessary change, it's important to verify that this specific commit contains only the intended fixes and does not introduce any regressions or unexpected behavior in the broader context of the engine. Subproject commit d7ff3afb90a5bb35bf2577fffc7d567c7ef064ed

[bong-water-water-bong]

Superseded by #104: this PR's llama.cpp commit is in the pin 96f6b89 (1bit-MONSTER/llama.cpp#4), with the MUL_MAT_ID route-stride fix and the FA value_stride range fix. All 7 HRX models pass serve_e2e.