perf(qwen35): optimize packed concurrent prefill on Strix Halo

[Graffioh]

Summary

Depends on #594. Optimizes Qwen3.6 concurrent prefill on Strix Halo while preserving per-request paged KV and recurrent state isolation.

Area	Optimization
Packed prefill	Fuses up to 8 FIFO prompt slices with live decode in one graph; batches projections, norms, FFN, and output work while running DeltaNet per ragged segment
Paged attention	Adds per-query causal positions so ragged prompt and decode rows share the same paged-attention traversal
Commit path	Runs the LM head only for prompt tails/decode rows; fixes M-RoPE axis-major staging; adds a 32x32 LDS concat transpose for short history and single-token decode
Adaptive planner	Uses a 2K mixed budget for short prompts and 4K at >=768 prompt tokens; keeps a 4K idle budget and 512-token quantum; all thresholds are environment-configurable
gfx1151 kernels	Enables hipBLASLt by default for wide concurrent shapes and tunes Q4_K MMQ tiles for narrow prefill/decode
Profiling	Adds ROCTX ranges around planning, graph build/upload/compute/readback, prefill, and decode phases

Validation

Check	Result
Release HIP/gfx1151 server build	Passed
Slot manager	152 checks passed
Batch planner	62 checks passed
Engine contract	14 checks passed
Diff/whitespace check	Passed

[cubic-dev-ai]

All reported issues were addressed across 67 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

1 issue found across 73 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="server/deps/llama.cpp/ggml/src/ggml-cpu/ops.cpp">

<violation number="1" location="server/deps/llama.cpp/ggml/src/ggml-cpu/ops.cpp:5008">
P1: Padded concurrent-decode rows crash on Metal and can write out of bounds on Vulkan because masked `SET_ROWS` semantics exist only in the CPU/CUDA paths. Add the same negative-index skip or reject/fallback behavior to every backend advertising `GGML_OP_SET_ROWS` before using this op-param.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 6 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 3 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 78 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>