6.2 KiB
Vulkan Chunked Gated Delta Net (GDN) — Performance & Development Notes
PR #20377 — First chunked parallel GDN implementation on any GPU shader backend.
Architecture
Three-stage chunked parallel decomposition (matches FLA/NVlabs reference implementations):
- Intra-chunk (
gated_delta_net_chunk_intra.comp) — Builds attention matrix A, computes W/U via WY representation. Outputs g_cumsum and total chunk decay. - Inter-chunk (
gated_delta_net_chunk_inter.comp) — Sequential across chunks, parallel across state columns. State update:S_next = exp(g_total) * S + K_gated^T @ v_corrected. - Output (
gated_delta_net_chunk_output_cm1.comp) — Coopmat GEMM kernel. ComputesA_decayed[64x64] @ vnew[64x128]using VK_KHR_cooperative_matrix (f16 inputs, f32 accumulation).
Chunk size: C=64 tokens. State dimensions: S_K=S_V=128. Pipeline: d128 non-KDA configs only.
Development History
Phase 1: Infrastructure (PR #20334, merged)
- Autoregressive GDN Vulkan shader — single-token sequential processing
- PP-512: 165 t/s, TG-128: 21.2 t/s on 890M (16 CU)
- 13/13 backend-ops tests
Phase 2: Graph-level chunked ops (PR #20340, merged)
- Chunked op decomposition at the GGML graph level
- Feeds autoregressive shader more efficiently
- PP-512: 165 → 220 t/s (+30.3%) — this gain is already in master
Phase 3: Vulkan chunked shaders (PR #20377, this PR)
- Three new compute shaders for intra/inter/output stages
- Initial scalar output kernel — functional but dispatch overhead made it slower than autoregressive on 16 CU
- Threshold gating: chunked path activates only when beneficial
Phase 4: Coopmat output kernel
- Replaced scalar output with VK_KHR_cooperative_matrix GEMM
- f16 shared memory for A_decayed and vnew, f32 accumulation via coopmat
- 4-phase architecture: QK^T via coopmat → decay mask → vnew staging → A_decayed @ vnew GEMM
- Numerically stable: direct
exp(g_i - g_j)per element (no factorization — factorized approach caused PPL regression to 20.06) - 16/16 backend-ops tests pass
Abandoned Approaches
- Factorized exp with g_max:
exp(g_max - gcum[j])amplified vnew, caused catastrophic cancellation. PPL 20.06 vs 13.46 baseline. - Scoped register split: Attempted to reduce VGPR pressure via scope boundaries. RADV compiler ignores scope for register allocation — no measurable difference.
Current Performance
Hardware: AMD Radeon 890M (RDNA3.5, 16 CU, 64KB LDS/CU, warp 64, KHR_coopmat) Model: Qwen3-Coder-Next-REAM Q4_K_M (60.33B params, 34.21 GiB)
Throughput (chunked coopmat, GDN_CHUNK_THRESHOLD=2)
| Test | t/s |
|---|---|
| PP-512 | 217.55 ± 1.41 |
| PP-1024 | 219.84 ± 4.00 |
| PP-2048 | 216.89 ± 1.94 |
| TG-128 | 21.76 ± 0.06 |
Autoregressive vs Chunked Comparison
| Test | Autoregressive | Chunked coopmat | Delta |
|---|---|---|---|
| PP-512 | 225.68 ± 3.00 | 217.55 ± 1.41 | -3.6% |
| PP-1024 | 229.63 ± 4.39 | 219.84 ± 4.00 | -4.3% |
| PP-2048 | 230.88 ± 1.44 | 216.89 ± 1.94 | -6.1% |
| TG-128 | 21.29 ± 0.03 | 21.76 ± 0.06 | +2.2% |
On 16 CU, autoregressive is 3.6-6.1% faster for PP due to lower dispatch overhead. Note autoregressive PP improves from 512→2048 while chunked stays flat — the gap widens on small hardware but the scaling characteristics favor chunked on wider hardware.
GDN kernel time comparison (PP-512):
- Autoregressive: 36 × 1,150 us = 41 ms (1.8% of total)
- Chunked (3 dispatches): 36 × 5,173 us = 186 ms (7.9% of total)
The chunked path's 3-dispatch overhead (intra + inter + output) accounts for the per-kernel cost difference, but end-to-end impact is only 3.6-6.1% since GDN is a small fraction of total wall time on this MoE model.
Perplexity Validation (WikiText-2, 299K tokens)
| Context | Chunked coopmat | f32 baseline | Delta |
|---|---|---|---|
| 512 (584 chunks) | 13.52 ± 0.11 | 13.46 | +0.06 |
| 4096 (73 chunks) | 10.18 ± 0.08 | 10.15 | +0.03 |
Both within error bars. Chunked coopmat path is numerically lossless.
Per-Kernel Timing (GGML_VK_PERF_LOGGER, PP-512)
GATED_DELTA_NET: 36 × 5173 us = 186 ms (7.9% of 2.35s total)
FLASH_ATTN_EXT: 12 × 783 us = 9.4 ms (0.4% of 2.35s total)
GDN is 7.9% of PP-512 wall time on this MoE-heavy model. MUL_MAT and MoE routing dominate the remaining 92%.
Scaling Analysis
Why flat PP scaling matters
PP-512/1024/2048 all within ±2 t/s. The chunked architecture processes fixed-size 64-token chunks — adding more tokens adds more chunks at constant cost each. Autoregressive dispatches scale linearly with token count (36 layers × N tokens = 36N sequential dispatches).
Why 16 CU doesn't show the crossover
- Chunked output kernel dispatches 3 shaders (intra + inter + output) vs 1 for autoregressive
- Each shader has launch overhead (~10-20 us) that dominates on small hardware
- The 64×64 @ 64×128 coopmat GEMM in the output kernel can't saturate 16 CUs
- On 40+ CU hardware (e.g., Strix Halo 8060S, discrete GPUs), the matmul-heavy chunked path has more headroom
GDN share grows with model density
On Qwen3-Next (384-expert MoE), GDN is only 8% of wall time. On GDN-dense architectures with fewer/no MoE layers, GDN's share would be 30-40%+, making the chunked optimization proportionally more impactful.
Key Files
| File | Purpose |
|---|---|
vulkan-shaders/gated_delta_net.comp |
Autoregressive kernel |
vulkan-shaders/gated_delta_net_chunk_intra.comp |
Intra-chunk (A matrix, WY) |
vulkan-shaders/gated_delta_net_chunk_inter.comp |
Inter-chunk (state update) |
vulkan-shaders/gated_delta_net_chunk_output.comp |
Original scalar output |
vulkan-shaders/gated_delta_net_chunk_output_cm1.comp |
Coopmat GEMM output |
ggml-vulkan.cpp:10409 |
GDN_CHUNK_THRESHOLD (dispatch gating) |
Test Commands
# Backend ops tests
./build/bin/test-backend-ops -b Vulkan0 -o GATED_DELTA_NET
# Benchmark
./build/bin/llama-bench -m <model> -ngl 99 -fa 1 -n 128 -p 512 --output md
# Perf logger
GGML_VK_PERF_LOGGER=1 ./build/bin/llama-bench -m <model> -ngl 99 -fa 1 -n 128 -p 512 -r 3 --output md
# Perplexity
./build/bin/llama-perplexity -m <model> -ngl 99 -fa 1 --ctx-size 4096 -f data/wikitext-2-raw/wiki.test.raw