cache-dit 0.2.10__py3-none-any.whl → 0.2.12__py3-none-any.whl

cache_dit/_version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.2.10'
-__version_tuple__ = version_tuple = (0, 2, 10)
+__version__ = version = '0.2.12'
+__version_tuple__ = version_tuple = (0, 2, 12)

cache_dit/compile/utils.py

@@ -39,6 +39,14 @@ def set_custom_compile_configs(
     # https://github.com/pytorch/pytorch/issues/153791
     torch._inductor.config.autotune_local_cache = False
 
+    if dist.is_initialized():
+        # Enable compute comm overlap
+        torch._inductor.config.reorder_for_compute_comm_overlap = True
+        # L20 64 GB/s, PCIe; A100/A800 NVLink 300 GB/s.
+        torch._inductor.config.intra_node_bw = (
+            64 if "L20" in torch.cuda.get_device_name() else 300
+        )
+
     FORCE_DISABLE_CUSTOM_COMPILE_CONFIG = (
         os.environ.get("CACHE_DIT_FORCE_DISABLE_CUSTOM_COMPILE_CONFIG", "0")
         == "1"
@@ -51,14 +59,6 @@ def set_custom_compile_configs(
         )
         return
 
-    if dist.is_initialized():
-        # Enable compute comm overlap
-        torch._inductor.config.reorder_for_compute_comm_overlap = True
-        # L20 64 GB/s, PCIe; A100/A800 NVLink 300 GB/s.
-        torch._inductor.config.intra_node_bw = (
-            64 if "L20" in torch.cuda.get_device_name() else 300
-        )
-
     # Below are default settings for torch.compile, you can change
     # them to your needs and test the performance
     torch._inductor.config.max_fusion_size = 64

cache_dit/metrics/metrics.py

@@ -334,19 +334,22 @@ compute_video_mse = partial(
 )
 
 
+METRICS_CHOICES = [
+    "psnr",
+    "ssim",
+    "mse",
+    "fid",
+    "all",
+]
+
+
 # Entrypoints
 def get_args():
+    global METRICS_CHOICES
     parser = argparse.ArgumentParser(
         description="CacheDiT's Metrics CLI",
         formatter_class=argparse.ArgumentDefaultsHelpFormatter,
     )
-    METRICS_CHOICES = [
-        "psnr",
-        "ssim",
-        "mse",
-        "fid",
-        "all",
-    ]
     parser.add_argument(
         "metrics",
         type=str,
@@ -383,13 +386,6 @@ def get_args():
         default=None,
         help="Path to predicted video or Dir to predicted videos",
     )
-    parser.add_argument(
-        "--enable-verbose",
-        "-verbose",
-        action="store_true",
-        default=False,
-        help="Show metrics progress verbose",
-    )
 
     # Image 1 vs N pattern
     parser.add_argument(
@@ -431,10 +427,52 @@ def get_args():
         default=1,
         help="Batch size for FID compute",
     )
+
+    # Verbose
+    parser.add_argument(
+        "--enable-verbose",
+        "-verbose",
+        action="store_true",
+        default=False,
+        help="Show metrics progress verbose",
+    )
+
+    # Format output
+    parser.add_argument(
+        "--summary",
+        "-s",
+        action="store_true",
+        default=False,
+        help="Summary the outupt metrics results",
+    )
+
+    # Addtional perf log
+    parser.add_argument(
+        "--perf-log",
+        "-plog",
+        type=str,
+        default=None,
+        help="Path to addtional perf log",
+    )
+    parser.add_argument(
+        "--perf-tag",
+        "-ptag",
+        type=str,
+        default=None,
+        help="Tag to parse perf time from perf log",
+    )
+    parser.add_argument(
+        "--extra-perf-tags",
+        "-extra-ptags",
+        nargs="+",
+        default=[],
+        help="Extra tags to parse perf time from perf log",
+    )
     return parser.parse_args()
 
 
 def entrypoint():
+    global METRICS_CHOICES
     args = get_args()
     logger.debug(args)
 
@@ -449,16 +487,19 @@ def entrypoint():
             batch_size=args.fid_batch_size,
         )
 
+    METRICS_META: dict[str, float] = {}
+
     # run one metric
     def _run_metric(
-        mertric: str,
+        metric: str,
         img_true: str = None,
         img_test: str = None,
         video_true: str = None,
         video_test: str = None,
     ) -> None:
         nonlocal FID
-        mertric = mertric.lower()
+        nonlocal METRICS_META
+        metric = metric.lower()
         if img_true is not None and img_test is not None:
             if any(
                 (
@@ -470,30 +511,30 @@ def entrypoint():
             # img_true and img_test can be files or dirs
             img_true_info = os.path.basename(img_true)
             img_test_info = os.path.basename(img_test)
-            if mertric == "psnr" or mertric == "all":
-                img_psnr, n = compute_psnr(img_true, img_test)
-                logger.info(
+
+            def _logging_msg(value: float, name, n: int):
+                if value is None or n is None:
+                    return
+                msg = (
                     f"{img_true_info} vs {img_test_info}, "
-                    f"Num: {n}, PSNR: {img_psnr}"
+                    f"Num: {n}, {name.upper()}: {value:.5f}"
                 )
-            if mertric == "ssim" or mertric == "all":
+                METRICS_META[msg] = value
+                logger.info(msg)
+
+            if metric == "psnr" or metric == "all":
+                img_psnr, n = compute_psnr(img_true, img_test)
+                _logging_msg(img_psnr, "psnr", n)
+            if metric == "ssim" or metric == "all":
                 img_ssim, n = compute_ssim(img_true, img_test)
-                logger.info(
-                    f"{img_true_info} vs {img_test_info}, "
-                    f"Num: {n}, SSIM: {img_ssim}"
-                )
-            if mertric == "mse" or mertric == "all":
+                _logging_msg(img_ssim, "ssim", n)
+            if metric == "mse" or metric == "all":
                 img_mse, n = compute_mse(img_true, img_test)
-                logger.info(
-                    f"{img_true_info} vs {img_test_info}, "
-                    f"Num: {n},  MSE: {img_mse}"
-                )
-            if mertric == "fid" or mertric == "all":
+                _logging_msg(img_mse, "mse", n)
+            if metric == "fid" or metric == "all":
                 img_fid, n = FID.compute_fid(img_true, img_test)
-                logger.info(
-                    f"{img_true_info} vs {img_test_info}, "
-                    f"Num: {n},  FID: {img_fid}"
-                )
+                _logging_msg(img_fid, "fid", n)
+
         if video_true is not None and video_test is not None:
             if any(
                 (
@@ -502,33 +543,33 @@ def entrypoint():
                 )
             ):
                 return
+
             # video_true and video_test can be files or dirs
             video_true_info = os.path.basename(video_true)
             video_test_info = os.path.basename(video_test)
-            if mertric == "psnr" or mertric == "all":
-                video_psnr, n = compute_video_psnr(video_true, video_test)
-                logger.info(
+
+            def _logging_msg(value: float, name, n: int):
+                if value is None or n is None:
+                    return
+                msg = (
                     f"{video_true_info} vs {video_test_info}, "
-                    f"Frames: {n}, PSNR: {video_psnr}"
+                    f"Frames: {n}, {name.upper()}: {value:.5f}"
                 )
-            if mertric == "ssim" or mertric == "all":
+                METRICS_META[msg] = value
+                logger.info(msg)
+
+            if metric == "psnr" or metric == "all":
+                video_psnr, n = compute_video_psnr(video_true, video_test)
+                _logging_msg(video_psnr, "psnr", n)
+            if metric == "ssim" or metric == "all":
                 video_ssim, n = compute_video_ssim(video_true, video_test)
-                logger.info(
-                    f"{video_true_info} vs {video_test_info}, "
-                    f"Frames: {n}, SSIM: {video_ssim}"
-                )
-            if mertric == "mse" or mertric == "all":
+                _logging_msg(video_ssim, "ssim", n)
+            if metric == "mse" or metric == "all":
                 video_mse, n = compute_video_mse(video_true, video_test)
-                logger.info(
-                    f"{video_true_info} vs {video_test_info}, "
-                    f"Frames: {n},  MSE: {video_mse}"
-                )
-            if mertric == "fid" or mertric == "all":
+                _logging_msg(video_mse, "mse", n)
+            if metric == "fid" or metric == "all":
                 video_fid, n = FID.compute_video_fid(video_true, video_test)
-                logger.info(
-                    f"{video_true_info} vs {video_test_info}, "
-                    f"Frames: {n},  FID: {video_fid}"
-                )
+                _logging_msg(video_fid, "fid", n)
 
     # run selected metrics
     if not DISABLE_VERBOSE:
@@ -574,7 +615,7 @@ def entrypoint():
         for metric in args.metrics:
             for img_test_dir in directories:
                 _run_metric(
-                    mertric=metric,
+                    metric=metric,
                     img_true=args.ref_img_dir,
                     img_test=img_test_dir,
                 )
@@ -619,7 +660,7 @@ def entrypoint():
         for metric in args.metrics:
             for video_test in video_source_selected:
                 _run_metric(
-                    mertric=metric,
+                    metric=metric,
                     video_true=args.ref_video,
                     video_test=video_test,
                 )
@@ -627,13 +668,169 @@ def entrypoint():
     else:
         for metric in args.metrics:
             _run_metric(
-                mertric=metric,
+                metric=metric,
                 img_true=args.img_true,
                 img_test=args.img_test,
                 video_true=args.video_true,
                 video_test=args.video_test,
             )
 
+    if args.summary:
+
+        def _fetch_perf():
+            if args.perf_log is None or args.perf_tag is None:
+                return []
+            if not os.path.exists(args.perf_log):
+                return []
+            perf_texts = []
+            with open(args.perf_log, "r") as file:
+                perf_lines = file.readlines()
+                for line in perf_lines:
+                    line = line.strip()
+                    if args.perf_tag.lower() in line.lower():
+                        if len(args.extra_perf_tags) == 0:
+                            perf_texts.append(line)
+                        else:
+                            has_all_extra_tag = True
+                            for ext_tag in args.extra_perf_tags:
+                                if ext_tag.lower() not in line.lower():
+                                    has_all_extra_tag = False
+                                    break
+                            if has_all_extra_tag:
+                                perf_texts.append(line)
+            return perf_texts
+
+        PERF_TEXTS: list[str] = _fetch_perf()
+
+        def _parse_value(
+            text: str,
+            tag: str = "Num",
+        ) -> float | None:
+            import re
+
+            escaped_tag = re.escape(tag)
+            processed_tag = escaped_tag.replace(r"\ ", r"\s+")
+
+            pattern = re.compile(
+                rf"{processed_tag}:\s*(\d+\.?\d*)\D*", re.IGNORECASE
+            )
+
+            match = pattern.search(text)
+
+            if not match:
+                return None
+
+            value_str = match.group(1)
+            try:
+                if tag.lower() in METRICS_CHOICES:
+                    return float(value_str)
+                if args.perf_tag is not None:
+                    if tag.lower() == args.perf_tag.lower():
+                        return float(value_str)
+                return int(value_str)
+            except ValueError:
+                return None
+
+        def _parse_perf(
+            compare_tag: str,
+        ) -> float | None:
+            nonlocal PERF_TEXTS
+            perf_times = []
+            for line in PERF_TEXTS:
+                if compare_tag in line:
+                    perf_time = _parse_value(line, args.perf_tag)
+                    if perf_time is not None:
+                        perf_times.append(perf_time)
+            if len(perf_times) == 0:
+                return None
+            return sum(perf_times) / len(perf_times)
+
+        def _format_item(
+            key: str,
+            metric: str,
+            value: float,
+            max_key_len: int,
+        ):
+            nonlocal PERF_TEXTS
+            # U1-Q0-C0-NONE vs U4-Q1-C1-NONE
+            header = key.split(",")[0].strip()
+            compare_tag = header.split("vs")[1].strip()  # U4-Q1-C1-NONE
+            has_perf_texts = len(PERF_TEXTS) > 0
+            format_str = ""
+            # Num / Frames
+            if n := _parse_value(key, "Num"):
+                if not has_perf_texts:
+                    format_str = (
+                        f"{header:<{max_key_len}}  Num: {n}  "
+                        f"{metric.upper()}: {value:<7.4f}"
+                    )
+                else:
+                    perf_time = _parse_perf(compare_tag)
+                    perf_time = f"{perf_time:<.2f}" if perf_time else None
+                    format_str = (
+                        f"{header:<{max_key_len}}  Num: {n}  "
+                        f"{metric.upper()}: {value:<7.4f}  "
+                        f"Perf: {perf_time}"
+                    )
+            elif n := _parse_value(key, "Frames"):
+                if not has_perf_texts:
+                    format_str = (
+                        f"{header:<{max_key_len}}  Frames: {n}  "
+                        f"{metric.upper()}: {value:<7.4f}"
+                    )
+                else:
+                    perf_time = _parse_perf(compare_tag)
+                    perf_time = f"{perf_time:<.2f}" if perf_time else None
+                    format_str = (
+                        f"{header:<{max_key_len}}  Frames: {n}  "
+                        f"{metric.upper()}: {value:<7.4f}  "
+                        f"Perf: {perf_time}"
+                    )
+            else:
+                raise ValueError("Num or Frames can not be NoneType.")
+
+            return format_str
+
+        selected_metrics = args.metrics
+        if "all" in selected_metrics:
+            selected_metrics = METRICS_CHOICES.copy()
+            selected_metrics.remove("all")
+
+        for metric in selected_metrics:
+            selected_items = {}
+            for key in METRICS_META.keys():
+                if metric.upper() in key or metric.lower() in key:
+                    selected_items[key] = METRICS_META[key]
+
+            reverse = True if metric.lower() in ["psnr", "ssim"] else False
+            sorted_items = sorted(
+                selected_items.items(), key=lambda x: x[1], reverse=reverse
+            )
+            selected_keys = [
+                key.split(",")[0].strip() for key in selected_items.keys()
+            ]
+            max_key_len = max(len(key) for key in selected_keys)
+
+            format_strs = []
+            for key, value in sorted_items:
+                format_strs.append(
+                    _format_item(key, metric, value, max_key_len)
+                )
+
+            format_len = max(len(format_str) for format_str in format_strs)
+
+            res_len = format_len - len(f"Summary: {metric.upper()}")
+            left_len = res_len // 2
+            right_len = res_len - left_len
+            print("-" * format_len)
+            print(
+                " " * left_len + f"Summary: {metric.upper()}" + " " * right_len
+            )
+            print("-" * format_len)
+            for format_str in format_strs:
+                print(format_str)
+            print("-" * format_len)
+
 
 if __name__ == "__main__":
     entrypoint()

{cache_dit-0.2.10.dist-info → cache_dit-0.2.12.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cache_dit
-Version: 0.2.10
+Version: 0.2.12
 Summary: 🤗 CacheDiT: A Training-free and Easy-to-use Cache Acceleration Toolbox for Diffusion Transformers
 Author: DefTruth, vipshop.com, etc.
 Maintainer: DefTruth, vipshop.com, etc
@@ -63,9 +63,8 @@ Dynamic: requires-python
 </div> 
 
 ## 🔥News🔥
-
-- [2025-07-13] **[🤗flux-faster](https://github.com/xlite-dev/flux-faster)** is released! A forked version of [huggingface/flux-fast](https://github.com/huggingface/flux-fast) that **makes flux-fast even faster** with **[cache-dit](https://github.com/vipshop/cache-dit)**, **3.3x** speedup on NVIDIA L20 while still maintaining **high precision**.
-  
+- [2025-07-18] 🎉First caching mechanism in **[🤗huggingface/flux-fast](https://github.com/huggingface/flux-fast)** with **[cache-dit](https://github.com/vipshop/cache-dit)**, also check the **[PR](https://github.com/huggingface/flux-fast/pull/13)**. 
+- [2025-07-13] **[🤗flux-faster](https://github.com/xlite-dev/flux-faster)** is released! A forked version of **[🤗huggingface/flux-fast](https://github.com/huggingface/flux-fast)** that **makes flux-fast even faster** with **[cache-dit](https://github.com/vipshop/cache-dit)**, **3.3x** speedup on NVIDIA L20.
 
 ## 🤗 Introduction 
 

{cache_dit-0.2.10.dist-info → cache_dit-0.2.12.dist-info}/RECORD

@@ -1,5 +1,5 @@
 cache_dit/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-cache_dit/_version.py,sha256=fMOyoyXAggjNTgl2YJ-8HW1bnjjDPiNACUsDoNufScI,513
+cache_dit/_version.py,sha256=7CFcHKqzy7OglwTX58ipGg1TD8MpDORqWvEBO3W1dHI,513
 cache_dit/logger.py,sha256=0zsu42hN-3-rgGC_C29ms1IvVpV4_b4_SwJCKSenxBE,4304
 cache_dit/primitives.py,sha256=A2iG9YLot3gOsZSPp-_gyjqjLgJvWQRx8aitD4JQ23Y,3877
 cache_dit/cache_factory/__init__.py,sha256=iYQwLwB_XLoYl0OB9unZGDbBtrYvZaLkOAmhGRwdW2E,191
@@ -31,17 +31,17 @@ cache_dit/cache_factory/first_block_cache/diffusers_adapters/hunyuan_video.py,sh
 cache_dit/cache_factory/first_block_cache/diffusers_adapters/mochi.py,sha256=lQTClo52OwPbNEE4jiBZQhfC7hbtYqnYIABp_vbm_dk,2363
 cache_dit/cache_factory/first_block_cache/diffusers_adapters/wan.py,sha256=dBNzHBECAuTTA1a7kLdvZL20YzaKTAS3iciVLzKKEWA,2638
 cache_dit/compile/__init__.py,sha256=DfMdPleFFGADXLsr7zXui8BTz_y9futY6rNmNdh9y7k,63
-cache_dit/compile/utils.py,sha256=OTvkwcezSrApZ2M1IMkYtkEmFbkfpTknhHMgoBApd6U,3786
+cache_dit/compile/utils.py,sha256=N4A55_8uIbEd-S4xyJPcrdKceI2MGM9BTIhJE63jyL4,3786
 cache_dit/custom_ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 cache_dit/custom_ops/triton_taylorseer.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 cache_dit/metrics/__init__.py,sha256=RaUhl5dieF40RqnizGzR30qoJJ9dyMUEADwgwMaMQrE,575
 cache_dit/metrics/config.py,sha256=ieOgD9ayz722RjVzk24bSIqS2D6o7TZjGk8KeXV-OLQ,551
 cache_dit/metrics/fid.py,sha256=9Ivtazl6mW0Bon2VXa-Ia5Xj2ewxRD3V1Qkd69zYM3Y,17066
 cache_dit/metrics/inception.py,sha256=pBVe2X6ylLPIXTG4-GWDM9DWnCviMJbJ45R3ulhktR0,12759
-cache_dit/metrics/metrics.py,sha256=O9a8qV6deQDWEoez7UZ_aqDLQ9rJXAJUMHGnJM7RUMs,19927
-cache_dit-0.2.10.dist-info/licenses/LICENSE,sha256=Dqb07Ik2dV41s9nIdMUbiRWEfDqo7-dQeRiY7kPO8PE,3769
-cache_dit-0.2.10.dist-info/METADATA,sha256=f5PF-lhexcLdB2HmBWETBEyg011eZOc99tlVI1lozYA,28002
-cache_dit-0.2.10.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-cache_dit-0.2.10.dist-info/entry_points.txt,sha256=FX2gysXaZx6NeK1iCLMcIdP8Q4_qikkIHtEmi3oWn8o,65
-cache_dit-0.2.10.dist-info/top_level.txt,sha256=ZJDydonLEhujzz0FOkVbO-BqfzO9d_VqRHmZU-3MOZo,10
-cache_dit-0.2.10.dist-info/RECORD,,
+cache_dit/metrics/metrics.py,sha256=1TTbfaj_-vdUfxopLnc5kVrXs5rMpAoSi8D0ItYdPu8,26439
+cache_dit-0.2.12.dist-info/licenses/LICENSE,sha256=Dqb07Ik2dV41s9nIdMUbiRWEfDqo7-dQeRiY7kPO8PE,3769
+cache_dit-0.2.12.dist-info/METADATA,sha256=-AIWGVOFsY-nhMkDeFErUFcELTWmza96-0IUN3od88A,28219
+cache_dit-0.2.12.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+cache_dit-0.2.12.dist-info/entry_points.txt,sha256=FX2gysXaZx6NeK1iCLMcIdP8Q4_qikkIHtEmi3oWn8o,65
+cache_dit-0.2.12.dist-info/top_level.txt,sha256=ZJDydonLEhujzz0FOkVbO-BqfzO9d_VqRHmZU-3MOZo,10
+cache_dit-0.2.12.dist-info/RECORD,,