op_config.csv

model_name,tag,repo
DeepSeek-R1,hq128tp1m1,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,128,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,128,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=128 S=32768/32768: {2*1*128*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq128tp1m4,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(4,128,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(4,128,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=4 HQ=128 S=32768/32768: {2*4*128*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq128tp1m16,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(16,128,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(16,128,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=16 HQ=128 S=32768/32768: {2*16*128*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq128tp1m0decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,128,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,128,8192,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=128 S=1/8192: {2*1*128*1*8192*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq128tp1m0prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,128,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,128,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=128 S=32768/32768: {2*1*128*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq16tp8m1,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,16,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,16,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=16 S=32768/32768: {2*1*16*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq16tp8m4,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(4,16,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(4,16,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=4 HQ=16 S=32768/32768: {2*4*16*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq16tp8m16,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(16,16,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(16,16,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=16 HQ=16 S=32768/32768: {2*16*16*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq16tp8m0decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,16,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,16,8192,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=16 S=1/8192: {2*1*16*1*8192*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq16tp8m0prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,16,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,16,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=16 S=32768/32768: {2*1*16*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
GLM-5,hq8tp8m0decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,8,1,256,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,8192,256,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=8 S=1/8192: {2*1*8*1*8192*2*256/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq8tp8m0prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,8,32768,256,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,32768,256,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=8 S=32768/32768: {2*1*8*32768*32768*2*256/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
Kimi-K2.5,hq64tp1m1,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,64,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,64,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=64 S=32768/32768: {2*1*64*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq64tp1m4,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(4,64,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(4,64,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=4 HQ=64 S=32768/32768: {2*4*64*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq64tp1m16,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(16,64,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(16,64,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=16 HQ=64 S=32768/32768: {2*16*64*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq64tp1m0decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,64,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,64,8192,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=64 S=1/8192: {2*1*64*1*8192*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq64tp1m0prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,64,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,64,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=64 S=32768/32768: {2*1*64*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq8tp8m1,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,8,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=8 S=32768/32768: {2*1*8*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq8tp8m4,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(4,8,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(4,8,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=4 HQ=8 S=32768/32768: {2*4*8*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq8tp8m16,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(16,8,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(16,8,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=16 HQ=8 S=32768/32768: {2*16*8*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq8tp8m0decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,8,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,8192,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=8 S=1/8192: {2*1*8*1*8192*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq8tp8m0prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,8,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=8 S=32768/32768: {2*1*8*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
MiniMax-M2.5,hq48tp1m0decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,48,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,8192,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=48 S=1/8192: {2*1*48*1*8192*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m1decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,48,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,4096,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=48 S=1/4096: {2*1*48*1*4096*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m16decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(16,48,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(16,8,4096,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=16 HQ=48 S=1/4096: {2*16*48*1*4096*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m64decode,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(64,48,1,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(64,8,4096,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=64 HQ=48 S=1/4096: {2*64*48*1*4096*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m0prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,48,32768,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,32768,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=48 S=32768/32768: {2*1*48*32768*32768*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m1prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(1,48,4096,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(1,8,4096,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=1 HQ=48 S=4096/4096: {2*1*48*4096*4096*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m16prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(16,48,4096,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(16,8,4096,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=16 HQ=48 S=4096/4096: {2*16*48*4096*4096*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""
,hq48tp1m64prefile,"python3 -c ""import torch,torch.nn.functional as F; q=torch.randn(64,48,4096,128,dtype=torch.bfloat16,device='cuda'); k=torch.randn(64,8,4096,128,dtype=torch.bfloat16,device='cuda'); v=k.clone(); k=k.repeat_interleave(6,dim=1); v=v.repeat_interleave(6,dim=1); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); import time; t=time.time(); [F.scaled_dot_product_attention(q,k,v) for _ in range(30)]; torch.cuda.synchronize(); ms=(time.time()-t)/30*1000; print(f'Attn B=64 HQ=48 S=4096/4096: {2*64*48*4096*4096*2*128/ms*1e-9:.1f} TFLOPS ({ms:.3f}ms)')"""