vulkan: Increase workgroup size for GLU, for performance

[jeffbolznv]

@CISC @0cc4m I noticed Vulkan perf was much worse for tg in #14158 due to the small workgroup size. This change restores the performance:

before:
Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m c:\models\glm-4-9b-chat-Q4_0.gguf -fa 1 -n 128 -p 512 --prio 1
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| chatglm 9B Q4_0                |   5.08 GiB |     9.40 B | Vulkan     |  99 |  1 |           pp512 |      3369.87 ± 10.71 |
| chatglm 9B Q4_0                |   5.08 GiB |     9.40 B | Vulkan     |  99 |  1 |           tg128 |         57.09 ± 0.21 |

build: ab46d11d (5752)

after:
Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m c:\models\glm-4-9b-chat-Q4_0.gguf -fa 1 -n 128 -p 512 --prio 1
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| chatglm 9B Q4_0                |   5.08 GiB |     9.40 B | Vulkan     |  99 |  1 |           pp512 |      3404.32 ± 11.38 |
| chatglm 9B Q4_0                |   5.08 GiB |     9.40 B | Vulkan     |  99 |  1 |           tg128 |         73.71 ± 0.24 |

build: 065b990f (5753)

[CISC]

I honestly don't know anything about the Vulkan backend, but if you say so I'm sure this is good. :)

[CISC]

Out of curiosity, is there a similar tg boost for models with split up/gate?

[jeffbolznv]

This was fixing a regression vs what's in master, so it's just recovering the performance we already had. I've only tested this one model.

[CISC]

This was fixing a regression vs what's in master, so it's just recovering the performance we already had. I've only tested this one model.

I understood, I guess what I was asking if you could check if there was a similar regression for split up/gate too?

[jeffbolznv]

There very likely was. Can you suggest a model to test?

[CISC]

Qwen3 or something?

[jeffbolznv]

Yes, there is a similar issue with Qwen3, which this mostly fixes. But it's still 1-2% slower. I think I need to change the shader to do one element per thread rather than a row per workgroup. I'll push another commit later today.

[jeffbolznv]

tg perf with qwen3 is now marginally faster than with master.