perf(BACKEND-ROCM): cooperate across Q8_K activation blocks

[VikashLoomba]

Row

BACKEND-ROCM

Before starting

• Issue QuantizeQ8KK is thread-per-superblock: 95.5us/call against llama.cpp's 1.6us, and 41% of the ROCm decode gap #1876 tracks the serial ROCm Q8_K activation quantizer. The owning specification lists it under ## Owed because the authenticated account cannot add its live Row: line.
• The recorded choice uses one pull request for specification and implementation.
• .agents/specs/rocm-q8k-cooperative-quantizer.md was committed before product code. It owns the byte oracle, production routes, architecture policy, acceptance gate, evidence, and remaining work.
• The implementation mirrors src/vt/cuda/cuda_quant_dot.cu::QuantizeQ8KPreqKernel and LaunchQuantizeQ8K. The CPU QuantizeRowQ8_K path supplies an independent byte oracle.

What changed

The ROCm Q8_K activation quantizer now cooperates across one 256-thread block instead of assigning one thread to each 256-element superblock. One shared launcher serves both dense and grouped production paths.

The encoder preserves all 292 bytes of BlockQ8_K, including signed first-occurrence ties. Kernel-resident test witnesses prove which arm executed instead of relying only on host dispatch counters.

Unset selection enables the cooperative arm only when the queue-device resolver returns gfx1100. Exact VT_ROCM_Q8K_BLOCK=0 keeps the permanent legacy arm, exact 1 forces the diagnostic candidate, and other values fail. Unset remains legacy for gfx1200, gfx1201, unknown architectures, and resolver failure.

Evidence

• Final reviewed head 0ee8722daac3199de6d395af099eff3d230c9fa5 has tree 31563a03b5a202a552fcd68c1fcce773295a962b. Fresh immutable review returned findings NONE and verdict PASS.
• The focused ROCm gate passed 3/3 cases and 3,561/3,561 assertions. It compared every byte with legacy GPU and CPU encoders across f32, f16, bf16, zero, random, signed-tie, padded, multi-row, and varied-superblock inputs.
• Dense and grouped public operations proved exclusive host selection and arm-distinct writes from inside the launched kernels for exact 0, exact 1, unset, and invalid values.
• Independent mutations disabled the accepted default, changed exact 0, bypassed grouped candidate execution, removed environment documentation, removed issue ownership, and injected an integration overlap. Each failed for its intended reason and restored byte-for-byte.
• Integrated and exact-parent full CTest each ran 700 tests with the same 32 failures and six skips. No integrated-only failure or skip appeared.
• The final unset real-checkpoint run produced 256 tokens with SHA-256 769bf8eebae5390db7b6aec5b9ab8e84caa9bc4124f659d77b7240f4494ed245.
• Hermetic full and staged preflights returned 0 and compiled 456/456 scoped translation units. They declared five argument-dependent checks skipped and explicitly did not call that result green.
• docs/ROCM.md, docs/USAGE.md, and docs/ENVIRONMENT.md document the architecture-scoped default and strict rollback control.

Speed claims

• The operator ran one discarded warmup and five same-binary pairs per arm under /home/vikash/gpu.lock on an RX 7900 XTX. The candidate won 5/5 pairs: median decode rose from 30.943 to 40.316 tok/s, a 31.377% median paired improvement with one-sided sign-test p=0.03125.
• Two rocprofv3 subtraction pairs per arm measured 129 quantizer calls per token. Median quantizer time fell from 7.8811 to 0.3426 ms/token, while total kernel time fell from 25.7287 to 18.1923 ms/token.
• The specification records the exact model, prompt, binaries, raw artifact hashes, profiler method, memory resolution, hardware mapping, contention checks, and rejected first campaign. This change adds no public benchmark ID, so docs/BENCHMARKS.md does not change.

Honest gaps

• gfx1200 and gfx1201 lack runtime acceptance. Their unset policy remains legacy, explicit 1 makes no support or speed claim there, and issue QuantizeQ8KK is thread-per-superblock: 95.5us/call against llama.cpp's 1.6us, and 41% of the ROCm decode gap #1876 remains open for that work.
• The dense checkpoint cannot exercise grouped engine routing. The focused production test independently proves the grouped launch with a kernel-resident witness.
• Current target main and the integrated head share the same 32 full-suite failures and six skips. The preflight's five argument-dependent checks were not run without their required inputs.
• The residual quantized GEMM gap remains owned by ROCm decode is 2.4x behind llama.cpp on a DENSE model (Ornith-1.5-9B, qwen35): the gap is not MoE-specific #1863 and is outside this change.

Addresses #1876.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:gpt-5 [codex]

[VikashLoomba]

CI classification for reviewed head 0ee8722daac3:

The three red jobs reproduce on target main and are not introduced by this six-path ROCm change.

• windows-msvc-cpu and windows-msvc-vulkan each fail the same three test_openai_api_server cases with rc=-1073740791: cases 56, 57, and 59. Main's scheduled run at ed5ecea2e has the same three failures in both jobs (CPU, Vulkan). The same signatures also predate this PR at main ancestor 7a1a98760 (scheduled run).
• sanitize-cpu (address,undefined) reports the same target-main failures. In particular, the 384-byte, four-allocation leak in test_glm_moe_dsa_schedule.cpp:680 is byte-for-byte the main signature; the remaining failed test names also match main's run at ed5ecea2e (main job, PR job).

The pull-request diff contains only the Q8_K specification, three documentation projections, src/vt/rocm/rocm_grouped_gemm.hip, and tests/vt/test_backend_cross_device.cpp. It does not change the Windows API server, GLM DSA scheduler, CPU backend, or sanitizer configuration. Existing issues #2403 and #2431 own the Windows aborts and GLM test leak, respectively.

All other executed PR jobs passed, including build-test-cpu, build-test-vulkan, build-newest-gcc, both container verification jobs, CUDA fat build, ARM64 CPU build, agent record, documentation, commit protocol, and PR size. The exact ROCm focused gate and real-checkpoint gate remain PASS as recorded in the body.