fix(ROCM): AttnQkNormRopeGate output-dtype dispatch -- the 0.8B GDN forward divergence root-caused and fixed (RED anchor turns GREEN)

[VikashLoomba]

fix(ROCM): AttnQkNormRopeGate dispatch keyed on the OUTPUT dtype — the 0.8B GDN forward divergence, root-caused and fixed (issue #41)

Restructured after review: this PR now carries ONLY the kernel fix + its teeth + the debug instrumentation, and lands BELOW the 0.8B gate (#559) so every commit is individually green on the gate host.

Root cause

The ROCm dispatcher for the fused full-attention preamble templated the kernel on the source dtype (qgate.dtype) where the CUDA lane (LaunchAttnPreambleOut) switches on the output dtype (q_out.dtype). The 0.8B bf16 model runs bf16 QKV-projection output + f32 q/k/gate outs (the f32-attention path; no FA-2 on ROCm) — silently mis-launched as all-bf16, writing bf16 bits through f32 out pointers.

Evidence chain (gfx1100, GPU-locked)

1. CPU backend == pinned oracle token-for-token; ROCm diverges at tok0 → ROCm-kernel-specific
2. Per-layer residual dumps (new VT_DUMP_ACT): layers 0–2 clean (≤9.6e-3 rms-rel), layer 3 (first full-attn) ignites to 2.1e-1
3. Sub-stage dumps (VT_DUMP_ACT_SUB): block_out 4.4e-1 with post_input_norm clean → inside FullAttnBlockPaged
4. Op dumps (VT_DUMP_ATTN): fa0_q rms-rel 1.196 with fa0_qkv/fa0_gate clean → the preamble's q/k path
5. VT_ROCM_ATTN_CPU_REF=1 unchanged → attention core exonerated

Fix + teeth

Dispatch on q_out.dtype; VT_CHECK bf16-out requires bf16-src. The cross-device case gains the bf16-src → f32-out combo at the real 0.8B dims (Hq=8, Hkv=2, Dh=256, rot=64) — mutation-proven: fails with the fix reverted, passes with it.

Gates (gfx1100, flock)

• test_backend_cross_device 19/19 (346 assertions)
• test_qwen3_paged_engine 16/16 unchanged
• e2e The capital of France is → Paris.\nThe capital of France is == pinned oracle == our CPU
• Full HIP ctest zero-delta vs base (box-baseline failures identical)

Also in this PR: three pre-existing hipcc/clang-only build breaks that blocked any full HIP build (same-line double CAPTURE ×2; -Wunused-const-variable on a CUDA-gated constant), and the env-gated dump instrumentation that found the bug (VT_DUMP_ACT/VT_DUMP_ACT_SUB/VT_DUMP_ATTN, off by default, allowlisted).

Current-main integration cutoff (2026-08-18)

Rebased from 053a18a42346e2610a8b56418ebd387c5e74a5c6 to fd836638d7a92bdbd755a7b73b9493f1ff9f6a0d on recorded current-main cutoff 65d6cdaed3e20e9bc70b4f9374fccafefefa7bd0. Exact output-dtype-fix commit replay. Fresh cutoff-head builds, focused gates, full repository preflight, and independent mutation review all pass.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: pi:kimi-k3 [pi]
Assisted-by: codex:gpt-5.6-sol [codex]
Assisted-by: pi:gpt-5.6-sol [pi]

[VikashLoomba]

Current-main rebase and fresh review

Rebased from c5b0b118 to ab41b731 on current-main pin f22c6cc8 and force-updated with an explicit lease.

• Fresh full cross-device gate: 20/20 cases, 352/352 assertions.
• The real Hq=8/Hkv=2/Dh=256/rotary=64 bf16-source/f32-output tooth remains.
• Fresh source-dtype-dispatch mutation rebuilt and failed as required: Q/K NMSE 1.51127/1.53128 versus 0.0005. Restored focused gate passed 53/53 assertions.
• Adjacent new-main documentation is preserved; topology remains gate(ROCM): M4 -- pinned vLLM-ROCm oracle on gfx1100 + ROCm device-golden lane in the Qwen3-0.6B near-tie gate #549 → this fix → gate(ROCM): M4 for Qwen3.5-0.8B GDN -- oracle-backed green near-tie gate #559.
• Repository preflight: All gates green.
• Fresh independent integration review: PASS; final tree clean.

[VikashLoomba]

Final integration cutoff

Final head: fd836638d7a92bdbd755a7b73b9493f1ff9f6a0d, exact replay on cutoff 65d6cdae.

• Cross-device: 20/20 cases, 352 assertions.
• Fresh source-dtype mutation RED: Q/K NMSE 1.51127/1.53128 versus 0.0005; restored focused arm 53/53 GREEN.
• New main Qwen3.5/runner changes remain intact; dispatch still keys on output dtype.
• Full preflight and independent cutoff review: PASS.

[localai-bot]

Landed via #559 as 7b89cf307, not superseded.

#559 is the tip of this stack and carried this branch's commit fd836638d (the AttnQkNormRopeGate output-dtype dispatch fix) verbatim. Verified on main: src/vt/rocm/rocm_gdn_fused.hip dispatches on q_out.dtype rather than the source dtype.

The tip was merged rather than each branch in turn because this repository is squash-only, and squash-merging a stack in order rewrites each ancestor and destroys the merge base of the branches above it. Authorship is preserved on the landed commit.

This one deserves a note: the evidence chain here, CPU backend matching the pinned oracle token-for-token, per-layer residual dumps localising the ignition to layer 3, then sub-stage and op dumps narrowing to fa0_q, then VT_ROCM_ATTN_CPU_REF=1 exonerating the attention core, is the kind of root-cause work that makes a fix reviewable. The bf16-src to f32-out cross-device case is mutation-proven, which is what the protocol asks for and rarely gets.