feat: TurboQuant KV-cache quantization for AMD ROCm (turbo3/turbo4)
Implements TurboQuant (Zandieh et al., ICLR 2026) KV-cache vector
quantization targeting AMD RDNA 4 (gfx1201, RX 9070 XT).
Algorithm: L2-normalize → FWHT(128) → Lloyd-Max scalar quantize → bitpack
Decode: unpack → codebook lookup → inverse FWHT → denormalize
Two new GGML types:
- GGML_TYPE_TURBO3_0: 3-bit, 3.5 bpw, MSE*d=0.034 (block_size=32, 14 bytes)
- GGML_TYPE_TURBO4_0: 4-bit, 4.5 bpw, MSE*d=0.009 (block_size=32, 18 bytes)
Architecture (pre-dequantize strategy):
- Write path: FWHT-aware set-rows kernels (128 threads, shared-mem FWHT)
- Read path: bulk dequantize turbo→f16 before standard Flash Attention
- Stride scaling preserves ggml_permute dim swaps (critical fix)
Performance (Qwen3-14B Q4_K_M, RX 9070 XT, 16 GB VRAM):
f16/f16: 1865 pp512, 54 tg128 (baseline)
q8_0/q8_0: 1694 pp512, 52 tg128
turbo4/turbo4: 1813 pp512, 49 tg128 (-3% pp, -9% tg, 72% less KV VRAM)
turbo3/turbo3: 1983 pp512, 49 tg128 (+6% pp, -9% tg, 78% less KV VRAM)
Usage: llama-cli -fa 1 --cache-type-k turbo4 --cache-type-v turbo4
Includes 7 CPU reference tests validating FWHT self-inverse, MSE against
paper values, bitpack determinism, and dequantize sanity.
Requires head_dim=128 (covers most current models including Llama, Qwen,
Mistral, Gemma). Guard added to KV cache init with clear error message.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
