| name: 07_w4a16_gemm | |
| display_name: "W4A16 Weight-only Quantized GEMM" | |
| precision: int4_bf16 | |
| regime: memory # decode-dominant; M=1 is bandwidth-bound on the int4 weight stream | |
| # Dense-equivalent FLOPs (matmul work, ignoring dequant arithmetic). | |
| flops_formula: "2 * M * N * K" | |
| # Bytes moved per call (memory roofline): | |
| # x: M*K*2 (bf16 activations, streamed in once) | |
| # w_q: (K/2)*N (packed int4, 0.5 B/elem) | |
| # scales: (K/128)*N*2 (bf16 scales) | |
| # zeros: (K/128)*N*2 (bf16 zero-points) | |
| # out: M*N*2 (bf16 store) | |
| bytes_formula: "M*K*2 + (K/2)*N + (K/128)*N*2 + (K/128)*N*2 + M*N*2" | |
| hardware: [RTX_PRO_6000] | |
| peak_tflops_key: bf16 | |
| peak_bandwidth_key: dram | |
| tolerance: | |
| bfloat16: 0.10 # group-quant adds noise on top of bf16 accumulator slop | |
| # Forbidden ops -- agent must write the unpack + GEMM themselves, not call a | |
| # vendor library that does both. | |
| forbidden: | |
| - "bitsandbytes.functional.dequantize_4bit" | |
| - "bitsandbytes.functional.gemv_4bit" | |
| - "marlin_kernel.gemm" | |
| - "torch.nn.functional.linear" | |
| sota: | |
| name: "bitsandbytes NF4 (gemv_4bit / dequantize_4bit + matmul)" | |
| url: "https://github.com/TimDettmers/bitsandbytes" | |
| function: "bitsandbytes.functional.gemv_4bit" | |
| notes: | | |
| Marlin (IST-DASLab) is the W4A16 SOTA on Ampere/Hopper but does not have | |
| SM120 (Blackwell consumer) kernels yet. GPTQ-Triton is unmaintained and | |
| does not target SM120. bitsandbytes 0.49.2 *does* run on SM120 -- it | |
| autotunes its CUDA kernels for compute capability 12.0 -- so we use its | |
| NF4 path (different quant scheme but same regime) as the SOTA reference | |
| line. Note that bnb's NF4 is symmetric/non-uniform; our reference uses | |
| AWQ-style asymmetric int4 with explicit zero-points, which is what the | |
| agent must implement. The SOTA line is informational only. | |
| deps: | |
| - "bitsandbytes>=0.49.2" | |
| num_correct_trials: 3 | |
| num_perf_trials: 50 | |