Vulkan LLM decode is ~11-13% slower than XNNPACK on Adreno 840 (Galaxy S26 Ultra), while prefill is ~2x faster. Qwen3, llama_main, cold-device medians, arms interleaved.
|
Vulkan (4w) |
XNNPACK (8da4w) |
| 0.6B decode |
98.0 tok/s |
109.0 tok/s |
| 1.7B decode |
41.3 tok/s |
47.6 tok/s |
| 0.6B prefill, 41-token prompt |
1568 tok/s |
818 tok/s |
At 0.6B, decode reads 335 MB of weights per token. The q4gsw GEMV is ~74% of GPU time at ~44.5 GB/s, and the GPU is ~100% busy.
#22941 (fused QK+softmax on the SDPA decode path) is the one change that helped: +2.5-3.9% decode.
Tried, no gain:
- 8da4w (
linear_dq8ca_q4gsw): 22% slower than 4w
- one command buffer per token: no change
- fusing SwiGLU (sigmoid + 2 muls): 0.998x
- tin GEMM at M=1 instead of the coop GEMV: 5.3x slower
- int3 weights: would beat XNNPACK, but +40% perplexity even with HQQ+AWQ
Possible directions: cut the non-GEMV work (SDPA ~12%, rms_norm ~5%, no-op view copies ~5%), or a sub-4-bit linear kernel paired with a quantizer that holds accuracy.
cc @SS-JIA @manuelcandales @digantdesai @cbilgin
Vulkan LLM decode is ~11-13% slower than XNNPACK on Adreno 840 (Galaxy S26 Ultra), while prefill is ~2x faster. Qwen3,
llama_main, cold-device medians, arms interleaved.At 0.6B, decode reads 335 MB of weights per token. The q4gsw GEMV is ~74% of GPU time at ~44.5 GB/s, and the GPU is ~100% busy.
#22941 (fused QK+softmax on the SDPA decode path) is the one change that helped: +2.5-3.9% decode.
Tried, no gain:
linear_dq8ca_q4gsw): 22% slower than 4wPossible directions: cut the non-GEMV work (SDPA ~12%, rms_norm ~5%, no-op view copies ~5%), or a sub-4-bit linear kernel paired with a quantizer that holds accuracy.
cc @SS-JIA @manuelcandales @digantdesai @cbilgin