cache-dit 0.2.36__py3-none-any.whl → 0.2.37__py3-none-any.whl

cache_dit/_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.2.36'
-__version_tuple__ = version_tuple = (0, 2, 36)
+__version__ = version = '0.2.37'
+__version_tuple__ = version_tuple = (0, 2, 37)
 
 __commit_id__ = commit_id = None

cache_dit/metrics/metrics.py

@@ -1,4 +1,5 @@
 import os
+import re
 import cv2
 import pathlib
 import argparse
@@ -160,25 +161,30 @@ def compute_dir_metric(
     # File
     if not os.path.isdir(image_true_dir) or not os.path.isdir(image_test_dir):
         return compute_file_func(image_true_dir, image_test_dir), 1
+
     # Dir
+    # compute dir metric
+    def natural_sort_key(filename):
+        match = re.search(r"(\d+)\D*$", filename)
+        return int(match.group(1)) if match else filename
+
     image_true_dir: pathlib.Path = pathlib.Path(image_true_dir)
-    image_true_files = sorted(
-        [
-            file
-            for ext in _IMAGE_EXTENSIONS
-            for file in image_true_dir.rglob("*.{}".format(ext))
-        ]
-    )
-    image_test_dir: pathlib.Path = pathlib.Path(image_test_dir)
-    image_test_files = sorted(
-        [
-            file
-            for ext in _IMAGE_EXTENSIONS
-            for file in image_test_dir.rglob("*.{}".format(ext))
-        ]
-    )
+    image_true_files = [
+        file
+        for ext in _IMAGE_EXTENSIONS
+        for file in image_true_dir.rglob("*.{}".format(ext))
+    ]
     image_true_files = [file.as_posix() for file in image_true_files]
+    image_true_files = sorted(image_true_files, key=natural_sort_key)
+
+    image_test_dir: pathlib.Path = pathlib.Path(image_test_dir)
+    image_test_files = [
+        file
+        for ext in _IMAGE_EXTENSIONS
+        for file in image_test_dir.rglob("*.{}".format(ext))
+    ]
     image_test_files = [file.as_posix() for file in image_test_files]
+    image_test_files = sorted(image_test_files, key=natural_sort_key)
 
     # select valid files
     image_true_files_selected = []
@@ -192,6 +198,7 @@ def compute_dir_metric(
         ):
             image_true_files_selected.append(selected_image_true)
             image_test_files_selected.append(selected_image_test)
+
     image_true_files = image_true_files_selected.copy()
     image_test_files = image_test_files_selected.copy()
     if len(image_true_files) == 0:
@@ -206,20 +213,22 @@ def compute_dir_metric(
 
     total_metric = 0.0
     valid_files = 0
+    total_files = 0
     for image_true, image_test in tqdm(
         zip(image_true_files, image_test_files),
         total=len(image_true_files),
         disable=DISABLE_VERBOSE,
     ):
         metric = compute_file_func(image_true, image_test)
-        if metric != float("inf"):
+        if metric != float("inf"):  # means no cache apply to image_test
             total_metric += metric
             valid_files += 1
+        total_files += 1
 
     if valid_files > 0:
         average_metric = total_metric / valid_files
         logger.debug(f"Average: {average_metric:.2f}")
-        return average_metric, valid_files
+        return average_metric, total_files
     else:
         logger.debug("No valid files to compare")
         return None, None
@@ -1217,6 +1226,9 @@ def entrypoint():
 
             if args.gen_markdown_table:
                 table = _format_table(format_strs, metric)
+                table = table.replace("Latency(s)(↑)", "SpeedUp(↑)")
+                table = table.replace("TFLOPs(↑)", "SpeedUp(↑)")
+                table = table.replace("FLOPs(↑)", "SpeedUp(↑)")
                 print("-" * format_len)
                 print(f"{table}")
             print("-" * format_len)

{cache_dit-0.2.36.dist-info → cache_dit-0.2.37.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cache_dit
-Version: 0.2.36
+Version: 0.2.37
 Summary: 🤗 A Unified and Training-free Cache Acceleration Toolbox for Diffusion Transformers
 Author: DefTruth, vipshop.com, etc.
 Maintainer: DefTruth, vipshop.com, etc
@@ -62,7 +62,8 @@ Dynamic: requires-python
  </div>
   <p align="center">
     <b><a href="#unified">📚Unified Cache APIs</a></b> | <a href="#forward-pattern-matching">📚Forward Pattern Matching</a> | <a href="#automatic-block-adapter">📚Automatic Block Adapter</a><br>
-    <a href="#hybird-forward-pattern">📚Hybrid Forward Pattern</a> | <a href="#dbcache">📚DBCache</a> | <a href="#taylorseer">📚Hybrid TaylorSeer</a> | <a href="#cfg">📚Cache CFG</a>
+    <a href="#hybird-forward-pattern">📚Hybrid Forward Pattern</a> | <a href="#dbcache">📚DBCache</a> | <a href="#taylorseer">📚Hybrid TaylorSeer</a> | <a href="#cfg">📚Cache CFG</a><br>
+    <a href="#benchmarks">📚Text2Image DrawBench</a> | <a href="#benchmarks">📚Text2Image Distillation DrawBench</a>
   </p>
   <p align="center">
     🎉Now, <b>cache-dit</b> covers <b>most</b> mainstream Diffusers' <b>DiT</b> Pipelines🎉<br>
@@ -275,11 +276,16 @@ Currently, **cache-dit** library supports almost **Any** Diffusion Transformers
 
 <div id="benchmarks"></div>
 
-Take FLUX.1-dev as an example. Here, only the results of some precision and performance benchmarks are presented. The test dataset is DrawBench. For a complete benchmark, please refer to [benchmarks](./bench/). **Device**: NVIDIA L20. **F**: Fn_compute_blocks, **B**: Bn_compute_blocks.
+cache-dit will support more mainstream Cache acceleration algorithms in the future. More benchmarks will be released, please stay tuned for update. Here, only the results of some precision and performance benchmarks are presented. The test dataset is **DrawBench**. For a complete benchmark, please refer to [📚Benchmarks](./bench/).
 
-| Config | Clip Score(↑) | ImageReward(↑) | PSNR(↑) | TFLOPs(↑) | SpeedUp(↑) |
+### 📚Text2Image DrawBench: FLUX.1-dev
+
+Comparisons between different FnBn compute block configurations show that **more compute blocks result in higher precision**. For example, the F8B0_W8MC0 configuration achieves the best Clip Score (33.007) and ImageReward (1.0333). **Device**: NVIDIA L20. **F**: Fn_compute_blocks, **B**: Bn_compute_blocks, 50 steps.
+
+
+| Config | Clip Score(↑) | ImageReward(↑) | PSNR(↑) | TFLOPs(↓) | SpeedUp(↑) |
 | --- | --- | --- | --- | --- | --- |
-| [**FLUX.1**-dev]: 50 steps | 32.9217 | 1.0412 | INF | 3726.87 | 1.00 |
+| [**FLUX.1**-dev]: 50 steps | 32.9217 | 1.0412 | INF | 3726.87 | 1.00x |
 | F8B0_W8MC0_R0.08 | 33.0070 | 1.0333 | 35.2008 | 2162.19 | 1.72x |
 | F8B0_W4MC0_R0.08 | 32.9871 | 1.0370 | 33.8317 | 2064.81 | 1.80x |
 | F4B0_W4MC2_R0.12 | 32.9718 | 1.0301 | 31.9394 | 1678.98 | 2.22x |
@@ -294,25 +300,57 @@ Take FLUX.1-dev as an example. Here, only the results of some precision and perf
 | F4B0_W4MC4_R0.12 | 32.8384 | 1.0065 | 31.5292 | 1400.08 | 2.66x |
 | F1B0_W4MC4_R0.12 | 32.8291 | 1.0181 | 32.9462 | 1401.61 | 2.66x |
 | F1B0_W4MC3_R0.12 | 32.8236 | 1.0166 | 33.0037 | 1457.62 | 2.56x |
+| F1B0_W4MC10_R1.0 | 32.3183 | 0.8796 | 29.6757 | 651.90 | 5.72x |
 
-The comparison between DBCache and algorithms such as Δ-DiT, Chipmunk, FORA, DuCa, TaylorSeer and FoCa is as follows. Now, in the comparison with a speedup ratio less than **3x**, cache-dit achieved the best accuracy. Please check [📚How to Reproduce?](./bench/) for more details.
+The comparison between **cache-dit: DBCache** and algorithms such as Δ-DiT, Chipmunk, FORA, DuCa, TaylorSeer and FoCa is as follows. Now, in the comparison with a speedup ratio less than **3x**, cache-dit achieved the best accuracy. Please check [📚How to Reproduce?](./bench/) for more details.
 
 | Method | TFLOPs(↓) | SpeedUp(↑) | ImageReward(↑) | Clip Score(↑) |
 | --- | --- | --- | --- | --- |
 | [**FLUX.1**-dev]: 50 steps | 3726.87 | 1.00× | 0.9898 | 32.404 |
 | [**FLUX.1**-dev]: 60% steps | 2231.70 | 1.67× | 0.9663 | 32.312 |
-| Δ-DiT (N=2) | 2480.01 | 1.50× | 0.9444 | 32.273 |
-| Δ-DiT (N=3) | 1686.76 | 2.21× | 0.8721 | 32.102 |
+| Δ-DiT(N=2) | 2480.01 | 1.50× | 0.9444 | 32.273 |
+| Δ-DiT(N=3) | 1686.76 | 2.21× | 0.8721 | 32.102 |
 | [**FLUX.1**-dev]: 34% steps | 1264.63 | 3.13× | 0.9453 | 32.114 |
 | Chipmunk | 1505.87 | 2.47× | 0.9936 | 32.776 |
 | FORA (N=3) | 1320.07 | 2.82× | 0.9776 | 32.266 |
-| **[DBCache (F=4, B=0)](https://github.com/vipshop/cache-dit)** | **1400.08** | **2.66×** | **1.0065** | **32.838** |
-| **[DBCache + TaylorSeer (F=4, B=0)](https://github.com/vipshop/cache-dit)** | **1388.30** | **2.68×** | **1.0287** | **32.914** |
+| **[DBCache(F=4,B=0,W=4,MC=4)](https://github.com/vipshop/cache-dit)** | **1400.08** | **2.66×** | **1.0065** | **32.838** |
 | DuCa(N=5) | 978.76 | 3.80× | 0.9955 | 32.241 |
-| TaylorSeer (N=4, O=2) | 1042.27 | 3.57× | 0.9857 | 32.413 |
-| **[FoCa (N=5) arxiv.2508.16211](https://arxiv.org/pdf/2508.16211)** | **893.54** | **4.16×** | **1.0029** | **32.948** |
-
-cache-dit will support more mainstream Cache acceleration algorithms in the future. More benchmarks will be released, please stay tuned for update. 
+| TaylorSeer(N=4,O=2) | 1042.27 | 3.57× | 0.9857 | 32.413 |
+| **[DBCache+TaylorSeer(F=1,B=0,O=1)](https://github.com/vipshop/cache-dit)** | **1153.05** | **3.23×** | **1.0221** | **32.819** |
+| **[FoCa(N=5) arxiv.2508.16211](https://arxiv.org/pdf/2508.16211)** | **893.54** | **4.16×** | **1.0029** | **32.948** |
+| [**FLUX.1**-dev]: 22% steps | 818.29 | 4.55× | 0.8183 | 31.772 |
+| FORA(N=4) | 967.91 | 3.84× | 0.9730 | 32.142 |
+| ToCa(N=8) | 784.54 | 4.74× | 0.9451 | 31.993 |
+| DuCa(N=7) | 760.14 | 4.89× | 0.9757 | 32.066 |
+| TeaCache(l=0.8) | 892.35 | 4.17× | 0.8683 | 31.704 |
+| **[DBCache(F=4,B=0,W=4,MC=10)](https://github.com/vipshop/cache-dit)** | 816.65 | 4.56x | 0.8245 | 32.191 |
+| TaylorSeer(N=5,O=2) | 893.54 | 4.16× | 0.9768 | 32.467 |
+| **[FoCa(N=7) arxiv.2508.16211](https://arxiv.org/pdf/2508.16211)** | **670.44** | **5.54×** | **0.9891** | **32.920** |
+| FORA(N=7) | 670.14 | 5.55× | 0.7418 | 31.519 |
+| ToCa(N=12) | 644.70 | 5.77× | 0.7155 | 31.808 |
+| DuCa(N=10) | 606.91 | 6.13× | 0.8382 | 31.759 |
+| TeaCache(l=1.2) | 669.27 | 5.56× | 0.7394 | 31.704 |
+| **[DBCache(F=1,B=0,W=4,MC=10)](https://github.com/vipshop/cache-dit)** | **651.90** | **5.72x** | 0.8796 | **32.318** |
+| TaylorSeer(N=7,O=2) | 670.44 | 5.54× | 0.9128 | 32.128 |
+| **[FoCa(N=8) arxiv.2508.16211](https://arxiv.org/pdf/2508.16211)** | **596.07** | **6.24×** | **0.9502** | **32.706** |
+
+NOTE: Except for DBCache, other performance data are referenced from the paper [FoCa, arxiv.2508.16211](https://arxiv.org/pdf/2508.16211).
+
+### 📚Text2Image Distillation DrawBench: Qwen-Image-Lightning
+
+Surprisingly, cache-dit: DBCache still works in the extremely few-step distill model. For example,  **Qwen-Image-Lightning w/ 4 steps**, with the F16B16 configuration, the PSNR is 34.8163, the Clip Score is 35.6109, and the ImageReward is 1.2614. It maintained a relatively high precision.
+
+| Config                     |  PSNR(↑)      | Clip Score(↑) | ImageReward(↑) | TFLOPs(↓)   | SpeedUp(↑) |
+|----------------------------|-----------|------------|--------------|----------|------------|
+| [**Lightning**]: 4 steps | INF       | 35.5797    | 1.2630       | 274.33   | 1.00x       |
+| F24B24_W2MC1_R0.8          | 36.3242   | 35.6224    | 1.2630       | 264.74   | 1.04x       |
+| F16B16_W2MC1_R0.8          | 34.8163   | 35.6109    | 1.2614       | 244.25   | 1.12x       |
+| F12B12_W2MC1_R0.8          | 33.8953   | 35.6535    | 1.2549       | 234.63   | 1.17x       |
+| F8B8_W2MC1_R0.8            | 33.1374   | 35.7284    | 1.2517       | 224.29   | 1.22x       |
+| F48B0_W2MC1_R0.8           | 30.0533   | 35.8483    | 1.1979       | 265.56   | 1.03x       |
+| F32B0_W2MC1_R0.8           | 29.6490   | 35.7684    | 1.2302       | 261.05   | 1.05x       |
+| F24B0_W2MC1_R0.8           | 29.6081   | 35.8599    | 1.1874       | 245.54   | 1.12x       |
+| F16B0_W2MC1_R0.8           | 29.4844   | 36.0810    | 1.1586       | 227.06   | 1.21x       |
 
 
 ## 🎉Unified Cache APIs
@@ -615,7 +653,7 @@ torch._dynamo.config.recompile_limit = 96  # default is 8
 torch._dynamo.config.accumulated_recompile_limit = 2048  # default is 256
 ```
 
-Please check [bench.py](./bench/bench.py) for more details.
+Please check [perf.py](./bench/perf.py) for more details.
 
 
 ## 🛠Metrics CLI

{cache_dit-0.2.36.dist-info → cache_dit-0.2.37.dist-info}/RECORD

@@ -1,5 +1,5 @@
 cache_dit/__init__.py,sha256=hzaexC1VQ0TxiWY6TJ1lTm-04e65WOTNHOfYryu1vFA,1284
-cache_dit/_version.py,sha256=y3W4kIzeljZ6sUAtO8hW9y1LA6HKsN9jWHCm6JjF5gw,706
+cache_dit/_version.py,sha256=jVUPlUOcnlQRBFP8i5PUv2oJntFMrKgk1rs1guuDZ34,706
 cache_dit/logger.py,sha256=0zsu42hN-3-rgGC_C29ms1IvVpV4_b4_SwJCKSenxBE,4304
 cache_dit/utils.py,sha256=nuHHr6NB286qE9u6klLNfhAVRMOGipihOhM8LRqznmU,10775
 cache_dit/cache_factory/.gitignore,sha256=5Cb-qT9wsTUoMJ7vACDF7ZcLpAXhi5v-xdcWSRit988,23
@@ -39,13 +39,13 @@ cache_dit/metrics/fid.py,sha256=ZM_FM0XERtpnkMUfphmw2aOdljrh1uba-pnYItu0q6M,1821
 cache_dit/metrics/image_reward.py,sha256=N8HalJo1T1js0dsNb2V1KRv4kIdcm3nhx7iOXJuqcns,5421
 cache_dit/metrics/inception.py,sha256=pBVe2X6ylLPIXTG4-GWDM9DWnCviMJbJ45R3ulhktR0,12759
 cache_dit/metrics/lpips.py,sha256=hrHrmdM-f2B4TKDs0xLqJO5JFaYcCjq2qNIR8oCrVkc,811
-cache_dit/metrics/metrics.py,sha256=RADSUUMYKBMkABsYFCEr_9PV8cDXLuxe2xuQ-mRBs4Y,39691
+cache_dit/metrics/metrics.py,sha256=7UV-H2NRbhfr6dvrXEzU97Zy-BSQ5zEfm9CKtaK4ldg,40231
 cache_dit/quantize/__init__.py,sha256=kWYoMAyZgBXu9BJlZjTQ0dRffW9GqeeY9_iTkXrb70A,59
 cache_dit/quantize/quantize_ao.py,sha256=Fx1KW4l3gdEkdrcAYtPoDW7WKBJWrs3glOHiEwW_TgE,6160
 cache_dit/quantize/quantize_interface.py,sha256=2s_R7xPSKuJeFpEGeLwRxnq_CqJcBG3a3lzyW5wh-UM,1241
-cache_dit-0.2.36.dist-info/licenses/LICENSE,sha256=Dqb07Ik2dV41s9nIdMUbiRWEfDqo7-dQeRiY7kPO8PE,3769
-cache_dit-0.2.36.dist-info/METADATA,sha256=2WIwYYezSUUk0rnIutttA2-s6_yTSCz1BwiAfO9cbQI,42706
-cache_dit-0.2.36.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-cache_dit-0.2.36.dist-info/entry_points.txt,sha256=FX2gysXaZx6NeK1iCLMcIdP8Q4_qikkIHtEmi3oWn8o,65
-cache_dit-0.2.36.dist-info/top_level.txt,sha256=ZJDydonLEhujzz0FOkVbO-BqfzO9d_VqRHmZU-3MOZo,10
-cache_dit-0.2.36.dist-info/RECORD,,
+cache_dit-0.2.37.dist-info/licenses/LICENSE,sha256=Dqb07Ik2dV41s9nIdMUbiRWEfDqo7-dQeRiY7kPO8PE,3769
+cache_dit-0.2.37.dist-info/METADATA,sha256=dLOxpSzGT1izGxxPdbFc7jDKtiSgl-XAl-JZXRkX138,45826
+cache_dit-0.2.37.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+cache_dit-0.2.37.dist-info/entry_points.txt,sha256=FX2gysXaZx6NeK1iCLMcIdP8Q4_qikkIHtEmi3oWn8o,65
+cache_dit-0.2.37.dist-info/top_level.txt,sha256=ZJDydonLEhujzz0FOkVbO-BqfzO9d_VqRHmZU-3MOZo,10
+cache_dit-0.2.37.dist-info/RECORD,,