[AaronStGeorge]

Overview

Add GGML_HRX_TRACE_COMPILE=1 to diagnose runtime kernel compilation during prefill and decode. It emits JSON records to stderr with graph execution and cache IDs, kernel keys and specialization parameters, compiler timing, failures, and unresolved-compilation waits. Asynchronous compiler events retain their originating execution ID. Graph-end records include zero compilation counts so repeated cycles can show cache reuse.

The caller chooses the workload and repetitions. Records are informational; backend graph IDs do not claim to identify token steps or prefill/decode phases. Existing asynchronous compilation and graph replay remain intact. Disabled tracing adds cached-state checks without timestamps, JSON construction, logging locks, or file I/O.

All changes are inside ggml/src/ggml-hrx. The README documents activation, events, timing limits, and JSON extraction.

Validation

Built llama-bench, hrx-backend-test, and hrx-loom-jit-test with HRX 82214dfb029975205d1c3a59164b678e2e21fc41 on a pinned gfx1100 GPU.

• Llama-3.2-1B-Instruct UD-Q4_K_XL, -p 0 -n 0 -pg 64,16 -b 64 -ub 64 -r 2 --no-warmup --progress: async and sync JIT both produced 68 backend graph executions and 43 initial compilation requests. Counts for the first four graphs were [1, 23, 1, 18]; all remaining graphs, including the entire second cycle, requested zero compilations.
• Parsed records to reconcile graph counts and match request/begin/end identities, timestamps, originating executions, and wait attribution. Sync and async compiled identical keys. Async recorded 16 waits; sync recorded none, as expected.
• Existing backend cases add_f32, scale_f32, and both same_uid_distinct_graph_reuses_graph_program variants: 4 passed with tracing disabled and 4 passed with tracing enabled. Disabled tracing emitted no trace records.
• Deliberately invalid GGML_HRX_LOOM_SANITIZER produced a matching failed compile-end record with compiler_started=false, the initialization error, and a failed graph-end record.
• Modified translation-unit syntax checks and git diff --check passed. The full backend suite was not run.

The existing compatibility floor is stale and is being corrected separately. Validation used the compatible HRX revision above; this PR does not change that floor or migrate kernel syntax for newer HRX revisions.

Requirements

• Reviewed the contributing guidance; this submission targets the AMD-Ecosystem fork.
• AI usage disclosure: YES. Codex implemented, reviewed, and tested the change and prepared this PR under user direction.