Skip to content

Commit

Permalink
fixes and improvements
Browse files Browse the repository at this point in the history 
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
  • Loading branch information
@HDCharles
HDCharles committed Aug 1, 2024
1 parent 2871e56 commit b8aaa96
Show file tree
Hide file tree
Showing 4 changed files with 80 additions and 24 deletions.
47 changes: 47 additions & 0 deletions torchao/_models/llama/benchmark_results.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,3 +29,50 @@
20240718151152, tok/s=100.29, mem/s=1325.29 GB/s, peak_mem=14.14 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240718151349, tok/s=166.08, mem/s=1099.70 GB/s, peak_mem= 8.95 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --quantization int8wo --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240718152147, tok/s=140.85, mem/s= 932.66 GB/s, peak_mem= 9.21 GB, model_size= 6.62 GB quant: int8wo, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --quantization int8wo --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8

20240731133002, tok/s=109.42, mem/s=1446.01 GB/s, peak_mem=13.90 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731135838, tok/s=102.85, mem/s=1359.17 GB/s, peak_mem=15.00 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731140259, tok/s=102.91, mem/s=1359.87 GB/s, peak_mem=15.00 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731140646, tok/s=101.19, mem/s=1337.23 GB/s, peak_mem=14.52 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731194813, tok/s=102.84, mem/s=1358.94 GB/s, peak_mem=15.00 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731195225, tok/s=103.14, mem/s=1362.92 GB/s, peak_mem=14.52 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731200747, tok/s=102.79, mem/s=1358.40 GB/s, peak_mem=15.00 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731201145, tok/s=103.09, mem/s=1362.33 GB/s, peak_mem=14.52 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8

20240731201438, tok/s=109.42, mem/s=1446.00 GB/s, peak_mem=13.88 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731201739, tok/s=102.58, mem/s=1355.51 GB/s, peak_mem=13.83 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731202121, tok/s=102.86, mem/s=1359.26 GB/s, peak_mem=15.00 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731202505, tok/s=103.44, mem/s=1366.91 GB/s, peak_mem=14.52 GB, model_size=13.21 GB quant: None, mod: Llama-2-7b-chat-hf, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Llama-2-7b-chat-hf/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8

20240731212356, tok/s= 95.36, mem/s=1431.41 GB/s, peak_mem=16.43 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731212822, tok/s= 93.92, mem/s=1409.76 GB/s, peak_mem=16.41 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731213330, tok/s= 89.84, mem/s=1348.49 GB/s, peak_mem=17.28 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731213900, tok/s= 88.34, mem/s=1326.01 GB/s, peak_mem=17.24 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731215214, tok/s= 80.92, mem/s=1214.62 GB/s, peak_mem=19.80 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8
20240731220805, tok/s= 78.54, mem/s=1178.91 GB/s, peak_mem=19.30 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8

20240731223923, tok/s= 95.52, mem/s=1433.68 GB/s, peak_mem=16.43 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731224400, tok/s= 90.39, mem/s=1356.73 GB/s, peak_mem=16.44 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240731224816, tok/s= 89.89, mem/s=1349.25 GB/s, peak_mem=17.26 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731225411, tok/s= 84.96, mem/s=1275.21 GB/s, peak_mem=17.48 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240731230612, tok/s= 80.91, mem/s=1214.45 GB/s, peak_mem=19.77 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8
20240731232130, tok/s= 69.10, mem/s=1037.25 GB/s, peak_mem=20.18 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8

20240801010740, tok/s= 95.45, mem/s=1432.64 GB/s, peak_mem=16.43 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240801011046, tok/s= 94.02, mem/s=1411.28 GB/s, peak_mem=16.41 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240801011513, tok/s= 89.96, mem/s=1350.32 GB/s, peak_mem=17.26 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240801011931, tok/s= 88.11, mem/s=1322.52 GB/s, peak_mem=17.20 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8

20240801013354, tok/s= 95.45, mem/s=1432.67 GB/s, peak_mem=16.43 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240801013812, tok/s= 92.15, mem/s=1383.16 GB/s, peak_mem=16.41 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240801025927, tok/s= 89.88, mem/s=1349.14 GB/s, peak_mem=17.26 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240801030347, tok/s= 87.32, mem/s=1310.69 GB/s, peak_mem=17.22 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240801031549, tok/s= 80.91, mem/s=1214.39 GB/s, peak_mem=19.77 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8
20240801033011, tok/s= 74.72, mem/s=1121.50 GB/s, peak_mem=19.34 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8

20240801093317, tok/s= 95.52, mem/s=1433.80 GB/s, peak_mem=16.43 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240801093529, tok/s= 92.36, mem/s=1386.35 GB/s, peak_mem=16.41 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 200 --top_k 200 --temperature 0.8
20240801093944, tok/s= 89.88, mem/s=1349.13 GB/s, peak_mem=17.26 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240801094415, tok/s= 87.20, mem/s=1308.88 GB/s, peak_mem=17.22 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 2048 --top_k 200 --temperature 0.8
20240801095615, tok/s= 80.87, mem/s=1213.82 GB/s, peak_mem=19.77 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: False, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8
20240801100912, tok/s= 74.65, mem/s=1120.41 GB/s, peak_mem=19.29 GB, model_size=15.01 GB quant: None, mod: Meta-Llama-3-8B, kv_quant: True, compile: True, compile_prefill: False, dtype: torch.bfloat16, device: cuda repro: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3-8B/model.pth --device cuda --precision torch.bfloat16 --kv_cache_quantization --compile --num_samples 5 --max_new_tokens 8192 --top_k 200 --temperature 0.8
Loading

0 comments on commit b8aaa96

Please sign in to comment.