Add native transparent PNG workflow and exact repeated-prompt encoder reuse with paired RGBA quality and timing evidence
371818e verified | import json,sys | |
| from pathlib import Path | |
| import torch,triton | |
| ROOT=Path(__file__).parent;sys.path.insert(0,str(ROOT.parent/'release')) | |
| from fp8_runtime import _quantize_rows | |
| def main(): | |
| rows=[] | |
| for m,k in [(4096,4096),(4096,12288),(16384,4096),(16384,12288)]: | |
| x=torch.randn(m,k,device='cuda',dtype=torch.bfloat16);s=torch.rand(k,device='cuda')+.05 | |
| q=torch.empty_like(x,dtype=torch.float8_e4m3fn);a=torch.empty(m,1,device='cuda') | |
| refq=torch.empty_like(q);refa=torch.empty_like(a) | |
| _quantize_rows[(m,)](x,s,refq,refa,k,triton.next_power_of_2(k),num_warps=8) | |
| for warps in [4,8,16,32]: | |
| def fn():_quantize_rows[(m,)](x,s,q,a,k,triton.next_power_of_2(k),num_warps=warps) | |
| fn();ms=triton.testing.do_bench(fn) | |
| row={'m':m,'k':k,'warps':warps,'ms':ms,'codes_equal':torch.equal(q.view(torch.uint8),refq.view(torch.uint8)),'scales_equal':torch.equal(a,refa)} | |
| rows.append(row);print(row,flush=True) | |
| (ROOT/'quantizer-trial.json').write_text(json.dumps(rows,indent=2)) | |
| main() | |