Row: GFX1100-TG200
Lands the ROCm keep-quant GEMM infrastructure: the kROCM provider for OpId::kMatmulBTQuant and OpId::kMatmulBTQuantGrouped. The provider does the dequant-in-kernel dot product on-device, keeping the weight in its block-quantized form through the GEMM rather than expanding to bf16 at load time.
VT_GGUF_KEEP_QUANT=0 restores load-time expansion. Gates on gfx1100 / ROCm 7.14.0: test_rocm_quant_dot 132,094 assertions green across all ten encodings (decode through prefill shapes, broadcast and per-row grouped arms over a poisoned output buffer).
This is the foundation PR for the GFX1100-TG200 decode campaign. Most lever PRs depend on the keep-quant GEMM infrastructure it introduces.
Row:
GFX1100-TG200Lands the ROCm keep-quant GEMM infrastructure: the
kROCMprovider forOpId::kMatmulBTQuantandOpId::kMatmulBTQuantGrouped. The provider does the dequant-in-kernel dot product on-device, keeping the weight in its block-quantized form through the GEMM rather than expanding to bf16 at load time.VT_GGUF_KEEP_QUANT=0restores load-time expansion. Gates on gfx1100 / ROCm 7.14.0:test_rocm_quant_dot132,094 assertions green across all ten encodings (decode through prefill shapes, broadcast and per-row grouped arms over a poisoned output buffer).This is the foundation PR for the GFX1100-TG200 decode campaign. Most lever PRs depend on the keep-quant GEMM infrastructure it introduces.