BACKEND-ROCM: ROCGdb CFA failure blocks causal-convolution instruction capture

Row: BACKEND-ROCM

The #2773 numerical-characterization acquisition cannot complete its instruction-step gate with the approved ROCm image. ROCGdb terminates with gdb/dwarf2/frame.c:1052: internal-error: Unknown CFA rule while preparing stepi at the first selected causal-convolution BF16 DOT2. This tracks the diagnostic capability gap under #2773; it does not change the underlying failing numerical test.

The operator reproduced this on 8 September 2026 with ROCGdb 16.3, rocm-rel-7.2-90, executable SHA256 8c239fef5ae0af68f478eee1f977f242fe38d8c0ffc2ec3c3fae9d1f4408e758, in image sha256:80aab4c182a1f3eeebe286173977e57fcaf10a049b41f475655b35d285de31dc. The unchanged oracle is vLLM e126687a9a828d513c01a07cd69f025f27d63280, Triton 3.7.1+f0b55c07, HIP 7.2.53211, on gfx1100 PCI 0000:03:00.0. Existing approved device/ptrace permissions were sufficient to stop and read the kernel. No runtime, compiler flags, registers, MODE, or PC were changed.

The unchanged collector selected test_causal_conv1d_varlen[4-False-4096-8-4-True-True-itype0]. Before the crash, the actual dispatch descriptor, live memory-URI source extent, and loaded instruction bytes were joined. The copied 9,816-byte ELF exactly matched SHA256 d2f70112304de9a32fbbddc0c530ad43ec4d0fbffda0290b64195d936e016518. The target was group (0,1,0), work-item (5,0,0), with original ABI registers and live sequence/chunk/cache-row/input/weight/bias words recorded.

At entry-relative PC +0x8ec (candidate ELF VA 0x21ec), v_dot2_bf16_bf16 v18.l, v18, v10, 0 had selected-lane sources 0x00003fbc and 0x00004062, MODE 0x3f0, and EXEC 0xffffffff. stepi then failed through dwarf2_frame_cache, compute_frame_id, set_step_info, prepare_one_step, and step_1. ROCGdb/container exit was 127. There is no after-instruction value, completed collector artifact, verified breakpoint restoration, or successful debugger preservation control. The partial acquisition is explicitly refused; the original numerical result remains unresolved by it.

A fresh plain control after the crash completed normally: the expected upstream AssertionError/pytest exit1 remained, and all 67 raw/f32 tensor files matched the accepted earlier capture. The original caches and failed debugger output are preserved separately.

Upstream has a related ROCGdb report #47 and a reported GDB fix a910478f65a596817fde3ad4ee727e41ec8e0513. That report is a source lead, not proof that its exact lower-level cause or fix applies to this binary.

Completion requires a source-bound, independently reviewed acquisition method that preserves this exact runtime/kernel and records actual before/after values, live instruction/coordinate linkage, normal breakpoint restoration, and complete repeated debugger/plain controls. Preserve this failed attempt. Do not silently replace the approved debugger/runtime or accept arithmetic predictions as measurements.

Start with the test_causal_conv1d_varlen[4-False-4096-8-4-True-True-itype0] collector described here and compare the ROCGdb failure with upstream ROCGdb report #47 and fix a910478f65a596817fde3ad4ee727e41ec8e0513. Done requires an independently reviewed acquisition on the exact runtime and kernel with before/after values, instruction and coordinate linkage, breakpoint restoration, and repeated debugger/plain controls.