feat(GFX1100-TG200): fuse silu-mul with Q8_K quant epilogue (VT_SILU_QUANT_FUSED) - #2890
Open
ghazni101 wants to merge 15 commits into
Open
feat(GFX1100-TG200): fuse silu-mul with Q8_K quant epilogue (VT_SILU_QUANT_FUSED)#2890ghazni101 wants to merge 15 commits into
ghazni101 wants to merge 15 commits into
Conversation
ghazni101
force-pushed
the
row/GFX1100-TG200-SILUQUANT
branch
from
September 4, 2026 07:46
7f91652 to
65fc569
Compare
This was referenced Sep 5, 2026
Collaborator
|
Held back with the rest of the I reviewed this change on its own and have no objection to it. Once #2790's base Landing today from this set: #2782 (with the grouped-Q8_0 repair), #2777 and |
ghazni101
force-pushed
the
row/GFX1100-TG200-SILUQUANT
branch
from
September 5, 2026 13:06
65fc569 to
bc6e532
Compare
Contributor
Author
|
Rebased onto the repaired stack tip 2ff6af4 (base b9f2ef4, the external-contributor landing branch). Gates green (check-env-doc, check-agent-record, check-rocm-dp4a-intrinsic); vllm-cli + test_rocm_quant_dot compile and link at -Werror in rocm-dev:10.0.0 on gfx1100; the default path (all arms off) is byte-identical to the staging baseline on Qwen3.5-4B Q4_K_M (canonical TG200 prompt, 256 tokens, greedy, seed 0). Engine A/B numbers measured today (one window, 5 reps, one load per arm): default 48.4 tok/s median; VT_GEMV_MMVQ=1 alone 53.5 (+10.0%); the byte-exact arm group (VT_NORM_QUANT_FUSED, VT_QUANT_Q8K_WARP) verified token-identical solo; VT_SKINNY_BF16 and the keep-quant arms (VT_GDN_ROWPERM_KEEP_QUANT / VT_GDN_COLPERM_KEEP_QUANT) and the association-change arms (VT_ATTN_DECODE_GQA4, VT_RMSNORM_ROW_COOP) move early near-tie tokens on the engine and are NOT engine-token-identical — flagged here so the acceptance terms name them; the full opt-in stack measured 72.5 tok/s median (+49.6%) with that documented divergence. This branch's head is now bc6e532. Ping for re-review. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true |
ghazni101
force-pushed
the
row/GFX1100-TG200-SILUQUANT
branch
from
September 5, 2026 16:43
bc6e532 to
3eca1d9
Compare
… kROCM The GGUF loader routes a block-typed weight to MatmulBTQuant whenever the running device has the provider, so registering these two ops lights up keep-quant compute on every ROCm board with no model-path change: the dense and grouped MoE towers stage once through ResidentWeight and dispatch to the new device GEMM. Coverage mirrors the CUDA sibling exactly — the ten Q8_K-family encodings plus a native Q8_0 arm. The integer dots are the portable scalar forms of the CPU reference bodies in the CPU accumulation order, because gfx1100 exposes no signed byte dot (v_dot4_i32_iu8 is unsigned-only; sdot4 needs a feature this target does not offer), and the gate is bit-exactness against the CPU tier at NMSE 1e-6 with the f64 dequant band at 5e-4. Unsupported dtypes throw naming the dtype instead of silently falling back to a host kernel that cannot follow device pointers; VT_GGUF_KEEP_QUANT=0 restores load-time expansion. Gates on gfx1100 / ROCm 7.14.0: test_rocm_quant_dot 132,094 assertions green across all ten encodings (decode through prefill shapes, broadcast and per-row grouped arms over a poisoned output buffer), focused ctest 'rocm|cross_device|quant' 20/21 with only the pre-existing MoeSiluMul bf16 exactness failure (mudler#1588) remaining, and an end-to-end Qwen3.5-0.8B Q4_K_M decode that is deterministic on device. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:ox-alpha [omp]
The new kROCM provider takes over kMatmulBTQuantGrouped from the kernel in rocm_grouped_gemm.hip and delegates Q4_K/Q5_K/Q6_K back to it, but not Q8_0. Q8_0 has no arm in rocm_quant_dot.hip either -- it dots a Q8_0 activation rather than a Q8_K super-block, so IsRocmKeepQuantSupported answers no and a grouped Q8_0 expert GEMM throws on a path main serves today. Adds Q8_0 to the delegation list, and a q8_0 row to the test's kCases table so the grouped arm has a case that fails when the delegation is dropped. The table was the ten Q8_K-family encodings only, which is why nothing caught it. Closes mudler#2927. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:claude-opus-5-1m [claude-code]
…CM provider" This reverts commit 82d99ea. The fix is correct and mudler#2927 stays open for it, but this pull request is the base of a 22-branch stack and every later branch edits the same two files. Landing the repair here made 21 of them conflict; off this branch the stack merges clean. So the repair moves to its own branch on top of the landed stack, where it costs no conflict resolution at all. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:claude-opus-5-1m [claude-code]
The new kROCM provider takes over kMatmulBTQuantGrouped from the kernel in rocm_grouped_gemm.hip and delegates Q4_K/Q5_K/Q6_K back to it, but not Q8_0. Q8_0 has no arm in rocm_quant_dot.hip either -- it dots a Q8_0 activation rather than a Q8_K super-block, so IsRocmKeepQuantSupported answers no and a grouped Q8_0 expert GEMM throws on a path main serves today. Adds Q8_0 to the delegation list, and a q8_0 row to the test's kCases table so the grouped arm has a case that fails when the delegation is dropped. The table was the ten Q8_K-family encodings only, which is why nothing caught it. Closes mudler#2927. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:claude-opus-5-1m [claude-code]
Issue mudler#1588 still lacks active cache-state evidence and a three-mode ROCm correctness gate. This spec fixes the post-write probes, dtype audit, tolerance policy, tests, review mutations, and hardware evidence before implementation starts. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:gpt-5.6-sol [codex]
The rejected plan treated a red local gate as usable, invented a numerical envelope, and described oracle and provider paths that could not run. Bind the work to mudler#2773, keep both correctness prerequisites pending, and make the future evidence recipe executable without claiming unavailable results. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:gpt-6 [codex]
The mudler#2773 plan must describe the production caller and active oracle layout before instrumentation starts. Correct BF16 selector normalization, name the existing Qwen3.5 path, and record its shared-seam debt in mudler#2923. Separate SD storage from DS dump order and cite the active CPU attention test with its unchanged tolerances. Runtime acceptance remains pending. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:gpt-6-astra [codex]
Adds the VT_GEMV_MMVQ=1 opt-in K-quant decode GEMV arm for MatmulBTQuant, bit-exact vs the CPU oracle. The arm folds activation quant into the MMVQ GEMV prologue (deleting the standalone QuantizeQ8KK launch) and widens the gate to engine dtypes (bf16/f16 activations, bf16/f32 outputs). Sub-levers: - lever-B1: VT_GEMV_MMVQ_FOLD_MAX makes the fold crossover tunable at runtime - lever-B2: VT_SKINNY_BF16=1 f32-out decode-skinny arm for GDN BA projections - repair: m-gates the whole dispatch and makes the GEMV bit-equal to baseline - repair-2: host-side dispatch-route counters + F1/F2 routing-witness gates - lever-B2 test: red-first f32-out decode-skinny gate, true-unset routing window Architecture: F1 moved the live MatmulBTQuantKernelRocm to rocm_quant_dot.hip (anonymous namespace, internal linkage). T4a's MMVQ arm lives in rocm_grouped_gemm.hip's version (external linkage, renamed to *Gdn). This PR adds delegation: rocm_quant_dot.hip forwards Q4_K/Q5_K/Q6_K calls to the Gdn version, preserving F1's IQ-type providers while activating T4a's MMVQ arm. The default path (VT_GEMV_MMVQ unset) is byte-unchanged from F1. The arm is opt-in and validated by test_rocm_quant_dot (6/6 cases, 719 assertions) and test_rocm_skinny_f32 (2/2 cases, 51 assertions). Token-identical to upstream baseline on Qwen3.5-4B Q4_K, 32-token greedy decode, seed 0. Rebased onto the external-contributor landing branch (staging tip b9f2ef4), which already carries F1 (mudler#2782) and its grouped-Q8_0 repair (mudler#2927). This version carries NONE of the removals the previous base commit 51f5222 made: the ten-row kCases table, kMaxNmseErr/nmse_ref_max and the three F1 provider cases are restored beside this arm's kKQuantCases (mudler#2938); Dp4a keeps the __ockl_sdot4 hardware dot (mudler#2939); the documented VT_ROCM_Q8K_BLOCK selector (SelectQ8KQuantArm/LaunchQ8KQuantizer), the mudler#2472 cooperative gfx1100 default (QuantizeQ8KCooperativeK) and the VT_ROCM_Q6K_SMALL_PRIVATE A/B arm are restored with their witness helpers; the shared bench-evidence file keeps lever B1's section 14 record, whose truncation this branch had carried. Depends on mudler#2782 (F1 keep-quant GEMM infra). FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM-5-2 [OMP] Assisted-by: AGENT:OMEN-ALPHA [OMP]
…Norm epilogue Lever-C adds an opt-in fused norm-quant epilogue (VT_NORM_QUANT_FUSED=1): RmsNormRowKernel emits the row's Q8_K blocks alongside its normal output, and MatmulBTQuant's K-quant branch skips the standalone QuantizeQ8KK when the consuming activation matches the producer token. Byte-identical to the standalone path by construction (shared QuantQ8KSBlock body). New files: - src/vt/rocm/rocm_act_quant.h: shared Q8_K quant-block body - src/vt/rocm/rocm_norm_quant_bridge.h: producer-consumer token contract Also fixes T4a routing counter placement (moved outside anonymous namespace for external linkage) and restores VT_GEMV_MMVQ_FOLD_MAX env var reading that was lost during cherry-pick conflict resolution. The default path (VT_NORM_QUANT_FUSED unset) is byte-unchanged. Validated by test_rocm_quant_dot (12/12 cases, 797 assertions). Token-identical to upstream baseline on Qwen3.5-4B Q4_K, 32-token greedy decode, seed 0. Depends on mudler#2782 (F1) and mudler#2790 (T4a). FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM-5-2 [OMP]
The standalone QuantizeQ8KK kernel used 1 thread per 256-element superblock, each doing a serial scan of 256 elements (~800 instructions). For decode (m=1, nsb=10) only 10 of 128 threads were active, and on wave32 each thread is its own wave, so the kernel took ~13.4 us/call = 540 us/tok (6.0% of wall time). The new QuantizeQ8KKWarpCoop kernel uses 8 threads per superblock (32 elements each). The amax scan is done per-chunk (ascending, ax > amax first-occurrence), then reduced across 8 threads via __shfl_xor_sync with lower-chunk-index tie-break — equivalent to a sequential scan of all 256 elements. The quantization (iscale = -127/mx, DNearestInt, clamp 127) and bsums are order-independent. Output is BYTE-IDENTICAL to the original QuantQ8KSBlock, asserted by the gate test (16/16, 839 assertions) under VT_QUANT_Q8K_WARP=1. For m=1, nsb=10: 1 block, 80/128 threads active (vs 10/128), 3 waves of ~100 instructions (vs 10 waves of ~800) = ~8x fewer wave-cycles. A/B on acceptance workload (Qwen3.5-4B Q4_K_M, 256 tokens, temp 0, seed 0): OFF median: 91.532 tok/s ON median: 93.417 tok/s +2.06%, 5/5 pairs ON>OFF, all 5 byte-identical (1039 bytes) Gated by VT_QUANT_Q8K_WARP (default OFF, read per-call). Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM-5-2 [OMP] FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM [omp]
…pKernel The fused Q8_K quant epilogue in RmsNormRowCoopKernel re-reads the normalized output from global memory (DLoadAct on orow) after Pass 3 stores it. On gfx1100 the 5 KB bf16 row (h=2560) competes with the weight and input in the 16 KB L1, so the re-read can miss to L2. T24 stores the normalized row to dynamic shared memory during Pass 3 (when the value is already in registers) and reads from LDS in the quant epilogue, eliminating the global re-read. The LDS buffer is h * sizeof(Tout) bytes (5 KB for bf16 h=2560), well within the 64 KB per-CU limit. Env gate VT_RMSNORM_LDS_QUANT (default ON) controls the optimization: set to 0 to revert to the global re-read path for A/B isolation. The gate is read per-call so captured graphs and in-process tests pick it up at dispatch time. Byte-identity: the LDS store uses the same conversion as Store (bf16 RNE for bf16 output, exact copy for f32), and DLoadAct reads the same bytes from LDS as from global. Gate test: 16/16 cases, 839 assertions, all passed. A/B measurement pending: the co-tenant 27B model holds the GPU VRAM, blocking the acceptance workload. The A/B script is staged at agent-artifacts/tg200-t24/ab-t24.sh for when the GPU is available. Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM-5-2 [OMP] FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM [omp]
…ctions The GDN layers attn_qkv (Q5_K, 24 tensors [2560,8192]) and attn_gate (Q4_K, 24 tensors [4096,2560]) were expanded to bf16 at load time because the V-head row reorder classified them as kTransformedWeight. The reorder is a ROW permutation — quantization blocks are along the K (column) dimension and are self-contained per row — so it is block-safe. T21 routes these tensors as kMatmulWeight to allow keep-quant, copies the blocks via OwnGgufQuantBlocks(mmap_src=nullptr), and applies ReorderVRows to the block bytes at load time. The forward pass already dispatches quantized nk=true weights through vt::MatmulBT, so no forward-pass change was needed. A/B: +3.9% (87.4 to 90.8 tok/s median, 5/5 pairs). Gate 16/16, 839 assertions. Output coherent but not byte-identical (Q5_K integer dot product vs bf16 float MAC). VT_GDN_ROWPERM_KEEP_QUANT=0 reverts to the old bf16 expansion path for A/B isolation. The improvement is less than the projected 14% because the Q5_K GEMV kernel has lower effective bandwidth on small grids (n=2560) than assumed, and wvSplitKSml is more efficient on these grids than projected. Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:glm-5-2 [omp] FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM [omp]
…t amax at 0 RmsNormRowCoopKernel's vec predicate constrained activation/residual widths but not the gamma type, while the vector body decodes the gamma as 8 bf16 halves per uint4. DispatchRmsNormWeight admits kF32/kF16 gammas independent of Tin (mudler#2492 decoupling), so with bf16 activation and an f32 or f16 gamma the vector path multiplied by garbage read from the wrong byte range. The predicate now requires Tw == __hip_bfloat16 exactly (sizeof==2 would still misread f16); the scalar path's Load() overloads decode per dtype and are unchanged. Also seeds the epilogue quantizer's mx/amax at 0 like the standalone QuantizeQ8KK: seeding from the thread's own element let a NaN poison d instead of producing the scalar contract's zero block.
ssm_out (out_proj) is Q5_K in the GGUF checkpoint but was expanded to bf16 at load time because the V-head column reorder (ReorderVCols) cuts across Q5_K 256-element block boundaries. T25 keeps the weight in tiled Q5_K order (no ReorderVCols) and permutes the 4096-element GEMV input from grouped to tiled order at runtime instead, cutting weight bandwidth ~4x (Q5_K ~5 MB vs bf16 20 MB per call). The permutation is a simple gather of 128-element groups within each of the 4096-element rows, gated by VT_GDN_COLPERM_KEEP_QUANT=1 (default OFF). A new out_proj_tiled flag on GdnLayerWeights distinguishes the tiled Q5_K path (needs input permutation) from the gdn_expand_nk bf16 path (already column-reordered, no permutation needed) — the nk flag alone conflates both. A/B (5 interleaved pairs, --max-tokens 256 --temperature 0 --seed 0): OFF median=90.930 tok/s, ON median=91.703 tok/s, +0.85%, 5/5 ON>OFF. Output coherent but NOT byte-identical (Q5_K vs bf16 weight precision). Gate test: 16/16, 839 assertions. The improvement is modest because the permutation kernel launch overhead (~13.4 us x 24 calls = ~322 us/tok) offsets most of the weight bandwidth savings (~368 us/tok). The net gain is ~46 us/tok. Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM-5-2 [OMP] FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM [omp]
The dense MLP's silu(gate)*up output feeds exactly one consumer: the K-quant ffn_down GEMV, which re-quantized the bf16 activation through a standalone QuantizeQ8KKWarpCoop launch (~8.9us x 32 calls per decode step, 6.5% of trace GPU time). The silu-mul producer now computes the Q8_K scratch in the same launch behind the adopted VT_NORM_QUANT_FUSED lever and records the shared producer/consumer bridge token, so the ffn_down dispatch skips that launch. VT_SILU_QUANT_FUSED=0 opts the silu site back out for a same-binary A/B. The new SiluMulQuantQ8KKWarpCoop kernel reuses the QuantizeQ8KKWarpCoop thread mapping and body verbatim (8 threads per super-block, ascending amax with lower-chunk-index tie-break, butterfly xor reduce). The byte-exactness argument: the split path stores silu(g)*up through __float2bfloat16 (SiluMulK::St) and the standalone quantizer loads those bits via DBF16ToF32; the fused kernel performs the identical store and consumes __bfloat162float of the same bf16, which is the same f32 the standalone quantizer reads. Asserted end-to-end: 256-token greedy output is byte-identical with the fusion on, and test_rocm_quant_dot (132,094 assertions) plus the backend/arch gates stay green. Measured on gfx1100, idle host (loadavg <= 0.68), acceptance workload (Qwen3.5-4B Q4_K_M, batch 1, greedy, 256 tokens, 17 adopted levers): median 91.387 tok/s fused vs 90.423 split over 5 runs each, every fused rep faster than every split rep (+1.07%, -0.117 ms/token). DSR ratchet holds at 32. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: AGENT:GLM [omp] Assisted-by: AGENT:OMEN-ALPHA [OMP]
ghazni101
force-pushed
the
row/GFX1100-TG200-SILUQUANT
branch
from
September 6, 2026 08:51
3eca1d9 to
75ce55b
Compare
Contributor
Author
|
Rebased onto FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true |
Contributor
|
For the gfx1100 campaign, the requested disposition remains close this experiment, or reduce it to historical evidence without product code. The reported median change for SiLU/quantization fusion is -0.08%, described here as noise. A source review with no objection does not establish a performance result that warrants shipping this opt-in path. At Please retain the original measurements, commands, revisions, and rejected outcome. Closing this PR does not establish a performance ceiling or close the broader gfx1100 gap. Revisit the lever when a current-main trace identifies a concrete remaining cost, followed by a same-binary A/B result and the matching correctness gate. This comment makes no new device measurement. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Row:
GFX1100-TG200Issue: #2885
Depends on: #2807 (T25)
Summary
Fuse the MLP silu(gate)*up activation with the Q8_K quantization epilogue on ROCm. When
VT_SILU_QUANT_FUSED=1,SiluMulQuantQ8KKWarpCoopproduces both the bf16 activation and the Q8_K scratch in a single kernel, deleting the separate ~8.9us quant launch.Benchmark
A/B interleaved, 5 pairs, Qwen3.5-4B Q4_K_M, 256 tokens, temp 0, seed 0,
--repeat 2(rep 1 warmup discarded):Noise — the fusion saves a launch but the benchmark is decode-dominated.
Token identity
PASS — identical output to T25 chain on
The capital of France iswith--max-tokens 32 --temperature 0 --seed 0.FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:GLM-5-2 [OMP]