Upload folder using huggingface_hub

quanto-exps/benchmark.py

@@ -26,8 +26,10 @@ def flush():
     torch.cuda.reset_max_memory_allocated()
     torch.cuda.reset_peak_memory_stats()
 
-def load_pipeline(ckpt_id, torch_dtype, qtype=None, exclude_layers=None, qte=False):
+def load_pipeline(ckpt_id, torch_dtype, qtype=None, exclude_layers=None, qte=False, fuse=False):
     pipe = DiffusionPipeline.from_pretrained(ckpt_id, torch_dtype=torch_dtype).to("cuda")
+    if fuse:
+        pipe.transformer.fuse_qkv_projections()
 
     if qtype:
         quantize(pipe.transformer, weights=qtype, exclude=exclude_layers)
@@ -76,13 +78,14 @@ if __name__ == "__main__":
     parser.add_argument("--torch_dtype", type=str, default="fp16", choices=list(TORCH_DTYPES.keys()))
     parser.add_argument("--qtype", type=str, default="none", choices=list(QTYPES.keys()))
     parser.add_argument("--qte", type=int, default=0, help="Quantize text encoder")
+    parser.add_argument("--fuse", type=int, default=0)
     parser.add_argument("--exclude_layers", metavar="N", type=str, nargs="*", default=None)
     args = parser.parse_args()
 
     flush()
 
     print(
-        f"Running with ckpt_id: {args.ckpt_id}, batch_size: {args.batch_size}, torch_dtype: {args.torch_dtype}, qtype: {args.qtype}, qte: {bool(args.qte)}"
+        f"Running with ckpt_id: {args.ckpt_id}, batch_size: {args.batch_size}, torch_dtype: {args.torch_dtype}, qtype: {args.qtype}, qte: {bool(args.qte)}, {args.exclude_layers=}, {args.fuse=}"
     )
     pipeline = load_pipeline(
         ckpt_id=args.ckpt_id,
@@ -90,6 +93,7 @@ if __name__ == "__main__":
         qtype=QTYPES[args.qtype],
         exclude_layers=args.exclude_layers,
         qte=args.qte,
+        fuse=bool(args.fuse),
     )
 
     for _ in range(WARM_UP_ITERS):
@@ -98,12 +102,12 @@ if __name__ == "__main__":
     time = benchmark_fn(run_inference, pipeline, args.batch_size)
     memory = bytes_to_giga_bytes(torch.cuda.max_memory_allocated())  # in GBs.
     print(
-        f"ckpt: {args.ckpt_id} batch_size: {args.batch_size}, qte: {args.qte}, "
+        f"ckpt: {args.ckpt_id} batch_size: {args.batch_size}, qte: {args.qte}, {args.exclude_layers=}"
         f"torch_dtype: {args.torch_dtype}, qtype: {args.qtype}  in {time} seconds and {memory} GBs."
     )
 
     ckpt_id = PREFIXES[args.ckpt_id]
-    img_name = f"ckpt@{ckpt_id}-bs@{args.batch_size}-dtype@{args.torch_dtype}-qtype@{args.qtype}-qte@{args.qte}"
+    img_name = f"ckpt@{ckpt_id}-bs@{args.batch_size}-dtype@{args.torch_dtype}-qtype@{args.qtype}-qte@{args.qte}-fuse@{args.fuse}"
     if args.exclude_layers:
         exclude_layers = "_".join(args.exclude_layers)
         img_name += f"-exclude@{exclude_layers}"
@@ -122,6 +126,7 @@ if __name__ == "__main__":
         qtype=args.qtype,
         qte=args.qte,
         exclude_layers=args.exclude_layers,
+        fuse=args.fuse,
     )
     info_file = f"{img_name}_info.json"
     with open(info_file, "w") as f:

quanto-exps/ckpt@auraflow-bs@1-dtype@bf16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "17.756", "time": "13.540", "dtype": "bf16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@1-dtype@bf16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "17.755", "time": "13.605", "dtype": "bf16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@1-dtype@fp16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "18.224", "time": "13.850", "dtype": "fp16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@1-dtype@fp16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "18.225", "time": "13.894", "dtype": "fp16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@4-dtype@bf16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "17.964", "time": "49.307", "dtype": "bf16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@4-dtype@bf16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "17.756", "time": "49.387", "dtype": "bf16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@4-dtype@fp16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "26.569", "time": "51.163", "dtype": "fp16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@auraflow-bs@4-dtype@fp16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "25.025", "time": "51.235", "dtype": "fp16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@1-dtype@bf16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "12.232", "time": "1.390", "dtype": "bf16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@1-dtype@bf16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "10.563", "time": "1.418", "dtype": "bf16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@1-dtype@fp16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "14.107", "time": "1.455", "dtype": "fp16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@1-dtype@fp16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "12.438", "time": "1.489", "dtype": "fp16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@4-dtype@bf16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "17.882", "time": "4.879", "dtype": "bf16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@4-dtype@bf16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "13.572", "time": "4.914", "dtype": "bf16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@4-dtype@fp16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "19.758", "time": "4.906", "dtype": "fp16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@pixart-bs@4-dtype@fp16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "13.573", "time": "4.946", "dtype": "fp16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@1-dtype@bf16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "15.314", "time": "2.522", "dtype": "bf16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@1-dtype@bf16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "15.264", "time": "2.579", "dtype": "bf16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@1-dtype@fp16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "17.187", "time": "2.516", "dtype": "fp16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@1-dtype@fp16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 1, "memory": "17.167", "time": "2.580", "dtype": "fp16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@4-dtype@bf16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "20.977", "time": "8.315", "dtype": "bf16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@4-dtype@bf16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "16.009", "time": "8.367", "dtype": "bf16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@4-dtype@fp16-qtype@int8-qte@0-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "22.847", "time": "8.375", "dtype": "fp16", "qtype": "int8", "qte": 0, "exclude_layers": null, "fuse": 1}

quanto-exps/ckpt@sd3-bs@4-dtype@fp16-qtype@int8-qte@1-fuse@1_info.json

@@ -0,0 +1 @@
+{"batch_size": 4, "memory": "17.171", "time": "8.477", "dtype": "fp16", "qtype": "int8", "qte": 1, "exclude_layers": null, "fuse": 1}

quanto-exps/launch_exps.sh

@@ -21,7 +21,7 @@ for ckpt_id in "${ckpt_ids[@]}"; do
           fi
 
           # Run the Python script with the current combination of arguments
-          python3 benchmark.py --ckpt_id $ckpt_id --batch_size $batch_size --torch_dtype $torch_dtype $q_arg --qte $te
+          python3 benchmark.py --ckpt_id $ckpt_id --batch_size $batch_size --torch_dtype $torch_dtype $q_arg --qte $te --fuse 1
         done
       done
     done