feat: Vulkan compute shader support for turbo3 (experimental)

[apollosenvy]

Summary

Vulkan backend support for turbo3 KV cache quantization. Experimental -- works on AMD 7900 XTX via RADV.

New files:

• dequant_turbo3_0.comp -- standalone dequant shader (3-bit index from 2-bit qs + 1-bit signs)
• dequant_funcs.glsl -- inline dequant/dequant4/get_dm for get_rows/mul_mat
• dequant_funcs_cm2.glsl -- cooperative matrix 2 FA path
• copy_to_quant.comp -- quantize with norm correction
• types.glsl -- block_turbo3_0 struct
• vulkan-shaders-gen.cpp -- turbo3_0 type registration
• ggml-vulkan.cpp -- pipeline creation + supports_op

Benchmark (AMD 7900 XTX, RADV, Vulkan)

Test	F16 KV	turbo3 KV	Ratio
pp128	748	264	35%
tg32	36.0	27.4	76%
tg128	33.3	26.4	79%

The tg numbers are already faster than ROCm HIP (25.2 t/s). The pp gap is from the standalone dequant path -- inline FA dequant would close it.

Status

• Quantize/dequant: working
• get_rows: working
• set_rows: working
• Flash attention: works via dequant-to-F16 path (not inline turbo3 FA)
• coopmat2 FA: shader compiles, untested on hardware

Marked experimental. Tested on RADV only.

🤖 Generated with Claude Code

[Titaniumtown]

Needs to be rebased to support gemma4 btw. At least I think so:

llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'gemma4'

[TheTom]

hey @apollosenvy . can you please rebase and test with HEAD. lots has changed and I apologize this fell through he crack.

[Titaniumtown]

I've been personally eyeing this PR because I have a small intel dgpu in my homeserver and want to use that for small llms.

I have been eyeing turboquant (and this PR) extensively :)

[apollosenvy]

Rebased onto current feature/turboquant-kv-cache HEAD (a4e8af44). Clean cherry-pick, no conflicts.

Build: Vulkan with GL_KHR_cooperative_matrix — compiles clean.

Benchmark (AMD 7900 XTX, RADV NAVI31, Vulkan):

Model	KV	pp128 (t/s)	tg128 (t/s)
Qwen3-8B Q4_K_M	f16 baseline	2780.81	120.85
Qwen3-8B Q4_K_M	q8_0-K / turbo3-V	393.85	44.52
Mistral-24B Q4_K_S	q8_0-K / turbo3-V	301.10	27.12

tg works well. The pp gap remains — this is the standalone dequant-to-F16 path, not inline turbo3 FA. Inline FA dequant would close it (noted in the original PR).

Quantize, dequant, get_rows, set_rows, and flash attention (via dequant path) all functional. Tested on RADV only.

[TheTom]

Built and tested on M5 Max (Metal4). No regression on Metal path:

Test	tok/s	Config
pp128	1359	Qwen3.5-35B-A3B ConfigI, q8_0-K/turbo3-V
tg32	40.2	same

All 7 changed files are Vulkan-only — zero changes to Metal, CUDA, or core ggml. Clean compile, no warnings.

Can't validate Vulkan shaders locally (no Vulkan GPU on Mac), but the code looks correct — PolarQuant centroids match, norm correction included, bit packing matches the Metal kernel layout.

Ship it.

[TheTom]

And thank you @apollosenvy

[apollosenvy]

Update: fused flash attention for turbo3 on Vulkan

Added inline turbo3 dequant directly in the FA compute shaders (scalar + coopmat1 paths). The standalone dequant-to-F16 indirection is gone for symmetric turbo3/turbo3.

Benchmark (AMD 7900 XTX, RADV NAVI31, Vulkan, turbo3/turbo3):

Model	Metric	Before (dequant path)	After (fused FA)	Gain
Qwen3-8B Q4_K_M	pp128	393.85 t/s	958.67 t/s	+143%
Qwen3-8B Q4_K_M	tg128	44.52 t/s	51.20 t/s	+15%
Mistral-24B Q4_K_S	pp128	301.10 t/s	481.24 t/s	+60%
Mistral-24B Q4_K_S	tg128	27.12 t/s	29.50 t/s	+9%

What changed:

• ggml-vulkan.cpp: CREATE_FA pipeline registration for GGML_TYPE_TURBO3_0 (scalar + cm1)
• vulkan-shaders-gen.cpp: emit turbo3_0 variants for both FA shader types
• flash_attn.comp / flash_attn_cm1.comp: enable FLASH_ATTN_DIRECT_TURBO3 define
• flash_attn_base.glsl: inline dequantize4() with Lloyd-Max centroid lookup

Known limitations:

• Asymmetric q8_0-K/turbo3-V not yet supported in fused path (falls back to dequant route). The Vulkan FA infrastructure requires K==V type; supporting asymmetric needs dual-dequant shader variants.
• Remaining gap to HIP/ROCm (99% of f16) is hardware-level: HIP uses v_dot2_f32_f16 inline ASM and sparse V skipping, neither accessible from Vulkan GLSL.

[Titaniumtown]

I found some issues on intel gpus I believe. I'm going to work on this with claude to see if I get anywhere with it.