File size: 19,115 Bytes
52baaab
 
 
 
c22e483
52baaab
 
2266161
7cae33f
 
9e8dd87
 
84ccd75
f2a58b9
 
84ccd75
 
68d414b
d6ed62a
f062324
68d414b
 
f062324
68d414b
1f1a0f8
 
3ce174d
1f1a0f8
2ec570a
 
 
 
68d414b
 
2ec570a
68d414b
f062324
4acff3b
9b4f8e6
 
5bca1b8
9b4f8e6
d237588
 
02469b3
5bca1b8
f807ff6
d237588
d350f56
 
 
 
 
dc3f6fc
8cbc120
 
 
02469b3
8cbc120
 
 
 
 
 
 
 
02469b3
 
8cbc120
 
 
 
 
 
 
 
 
 
 
 
02469b3
 
 
8cbc120
 
 
 
 
02469b3
8cbc120
 
 
 
 
 
02469b3
8cbc120
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c724a07
79a04cc
02469b3
 
 
3dc5d63
02469b3
 
79a04cc
02469b3
79a04cc
 
 
 
 
 
 
 
 
 
02469b3
 
 
 
 
 
 
 
 
 
79a04cc
 
d350f56
 
dc3f6fc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0a3f3cb
 
 
b5508c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0a3f3cb
b5508c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0a3f3cb
 
dc3f6fc
 
c04fadc
 
b1b6194
 
 
d350f56
fae9f22
 
1f1a0f8
 
d6ed62a
6917744
369d5ab
 
190d547
d66c08a
 
 
 
7cae33f
 
cac27a3
 
190d547
 
cac27a3
 
0523cf8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
---
license: apache-2.0
---

# Repos
https://github.com/mit-han-lab/deepcompressor

# Installation
https://github.com/mit-han-lab/deepcompressor/issues/56

https://github.com/nunchaku-tech/deepcompressor/issues/80

# Windows
https://learn.microsoft.com/en-us/windows/wsl/install

https://www.anaconda.com/docs/getting-started/miniconda/install

# Environment
python 3.12

cuda 12.8

torch 2.7

diffusers https://github.com/huggingface/diffusers

transformers 4.51

# Calibration

https://github.com/nunchaku-tech/deepcompressor/blob/main/examples/diffusion/README.md#step-2-calibration-dataset-preparation

# Quantization

https://github.com/nunchaku-tech/deepcompressor/blob/main/examples/diffusion/README.md#step-3-model-quantization

Model Path: https://github.com/nunchaku-tech/deepcompressor/issues/70#issuecomment-2788155233

Save model: `--save-model true` or `--save-model /PATH/TO/CHECKPOINT/DIR`

Example: `python -m deepcompressor.app.diffusion.ptq examples/diffusion/configs/model/flux.1-dev.yaml examples/diffusion/configs/svdquant/nvfp4.yaml`

Folder Structure 

- refer [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main)

- refer [black-forest-labs/FLUX.1-Kontext-dev](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev/tree/main)

---

# Blockers
1) NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

Potential fix: app.diffusion.pipeline.config.py
```python
    @staticmethod
    def _default_build(
        name: str, path: str, dtype: str | torch.dtype, device: str | torch.device, shift_activations: bool
    ) -> DiffusionPipeline:
        if not path:
            if name == "sdxl":
                path = "stabilityai/stable-diffusion-xl-base-1.0"
            elif name == "sdxl-turbo":
                path = "stabilityai/sdxl-turbo"
            elif name == "pixart-sigma":
                path = "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS"
            elif name == "flux.1-kontext-dev":
                path = "black-forest-labs/FLUX.1-Kontext-dev"
            elif name == "flux.1-dev":
                path = "black-forest-labs/FLUX.1-dev"
            elif name == "flux.1-canny-dev":
                path = "black-forest-labs/FLUX.1-Canny-dev"
            elif name == "flux.1-depth-dev":
                path = "black-forest-labs/FLUX.1-Depth-dev"
            elif name == "flux.1-fill-dev":
                path = "black-forest-labs/FLUX.1-Fill-dev"
            elif name == "flux.1-schnell":
                path = "black-forest-labs/FLUX.1-schnell"
            else:
                raise ValueError(f"Path for {name} is not specified.")
        if name in ["flux.1-kontext-dev"]:
            pipeline = FluxKontextPipeline.from_pretrained(path, torch_dtype=dtype)
        elif name in ["flux.1-canny-dev", "flux.1-depth-dev"]:
            pipeline = FluxControlPipeline.from_pretrained(path, torch_dtype=dtype)
        elif name == "flux.1-fill-dev":
            pipeline = FluxFillPipeline.from_pretrained(path, torch_dtype=dtype)
        elif name.startswith("sana-"):
            if dtype == torch.bfloat16:
                pipeline = SanaPipeline.from_pretrained(path, variant="bf16", torch_dtype=dtype, use_safetensors=True)
                pipeline.vae.to(dtype)
                pipeline.text_encoder.to(dtype)
            else:
                pipeline = SanaPipeline.from_pretrained(path, torch_dtype=dtype)
        else:
            pipeline = AutoPipelineForText2Image.from_pretrained(path, torch_dtype=dtype)

        # Debug output
        print(">>> DEVICE:", device)
        print(">>> PIPELINE TYPE:", type(pipeline))
    
        # Try to move each component using .to_empty()
        for name in ["unet", "transformer", "vae", "text_encoder"]:
            module = getattr(pipeline, name, None)
            if isinstance(module, torch.nn.Module):
                try:
                    print(f">>> Moving {name} to {device} using to_empty()")
                    module.to_empty(device)
                except Exception as e:
                    print(f">>> WARNING: {name}.to_empty({device}) failed: {e}")
                    try:
                        print(f">>> Falling back to {name}.to({device})")
                        module.to(device)
                    except Exception as ee:
                        print(f">>> ERROR: {name}.to({device}) also failed: {ee}")
    
        # Identify main model (for patching)
        model = getattr(pipeline, "unet", None) or getattr(pipeline, "transformer", None)
        if model is not None:
            replace_fused_linear_with_concat_linear(model)
            replace_up_block_conv_with_concat_conv(model)
            if shift_activations:
                shift_input_activations(model)
        else:
            print(">>> WARNING: No model (unet/transformer) found for patching")
    
        return pipeline
```

Debug Log
```
25-07-22 20:11:56 | I | === Start Evaluating ===
25-07-22 20:11:56 | I | * Building diffusion model pipeline
Loading pipeline components...:   0%|                                                             | 0/7 [00:00<?, ?it/s]
You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
Loading checkpoint shards: 100%|██████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 18.92it/s]
Loading pipeline components...: 100%|█████████████████████████████████████████████████████| 7/7 [00:00<00:00,  9.50it/s]
>>> DEVICE: cuda
>>> PIPELINE TYPE: <class 'diffusers.pipelines.flux.pipeline_flux_kontext.FluxKontextPipeline'>
>>> Moving transformer to cuda using to_empty()
>>> WARNING: transformer.to_empty(cuda) failed: Module.to_empty() takes 1 positional argument but 2 were given
>>> Falling back to transformer.to(cuda)
>>> ERROR: transformer.to(cuda) also failed: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.
>>> Moving vae to cuda using to_empty()
>>> WARNING: vae.to_empty(cuda) failed: Module.to_empty() takes 1 positional argument but 2 were given
>>> Falling back to vae.to(cuda)
>>> Moving text_encoder to cuda using to_empty()
>>> WARNING: text_encoder.to_empty(cuda) failed: Module.to_empty() takes 1 positional argument but 2 were given
>>> Falling back to text_encoder.to(cuda)
25-07-22 20:11:59 | I |   Replacing fused Linear with ConcatLinear.
25-07-22 20:11:59 | I |     + Replacing fused Linear in single_transformer_blocks.0 with ConcatLinear.
25-07-22 20:11:59 | I |       - in_features = 3072/15360
25-07-22 20:11:59 | I |       - out_features = 3072
25-07-22 20:11:59 | I |     + Replacing fused Linear in single_transformer_blocks.1 with ConcatLinear.
25-07-22 20:11:59 | I |       - in_features = 3072/15360
25-07-22 20:11:59 | I |       - out_features = 3072
25-07-22 20:11:59 | I |     + Replacing fused Linear in single_transformer_blocks.2 with ConcatLinear.
25-07-22 20:11:59 | I |       - in_features = 3072/15360
25-07-22 20:11:59 | I |       - out_features = 3072
```

2) KeyError: <class 'diffusers.models.transformers.transformer_flux.FluxAttention'>

Potential fix: app.diffusion.nn.struct.py
```python
    @staticmethod
    def _default_construct(
        module: Attention,
        /,
        parent: tp.Optional["DiffusionTransformerBlockStruct"] = None,
        fname: str = "",
        rname: str = "",
        rkey: str = "",
        idx: int = 0,
        **kwargs,
    ) -> "DiffusionAttentionStruct":
        if isinstance(module, FluxAttention):  
            # FluxAttention has different attribute names than standard attention  
            with_rope = True  
            num_query_heads = module.heads  # FluxAttention uses 'heads', not 'num_heads'  
            num_key_value_heads = module.heads  # FLUX typically uses same for q/k/v  
              
            # FluxAttention doesn't have 'to_out', but may have other output projections  
            # Check what output projection attributes actually exist  
            o_proj = None  
            o_proj_rname = ""  
              
            # Try to find the correct output projection  
            if hasattr(module, 'to_out') and module.to_out is not None:  
                o_proj = module.to_out[0] if isinstance(module.to_out, (list, tuple)) else module.to_out  
                o_proj_rname = "to_out.0" if isinstance(module.to_out, (list, tuple)) else "to_out"  
            elif hasattr(module, 'to_add_out'):  
                o_proj = module.to_add_out  
                o_proj_rname = "to_add_out"  
              
            q_proj, k_proj, v_proj = module.to_q, module.to_k, module.to_v  
            q_proj_rname, k_proj_rname, v_proj_rname = "to_q", "to_k", "to_v"  
            q, k, v = module.to_q, module.to_k, module.to_v  
            q_rname, k_rname, v_rname = "to_q", "to_k", "to_v"  
              
            # Handle the add_* projections that FluxAttention has  
            add_q_proj = getattr(module, "add_q_proj", None)  
            add_k_proj = getattr(module, "add_k_proj", None)   
            add_v_proj = getattr(module, "add_v_proj", None)  
            add_o_proj = getattr(module, "to_add_out", None)  
            add_q_proj_rname = "add_q_proj" if add_q_proj else ""  
            add_k_proj_rname = "add_k_proj" if add_k_proj else ""  
            add_v_proj_rname = "add_v_proj" if add_v_proj else ""  
            add_o_proj_rname = "to_add_out" if add_o_proj else ""  
              
            kwargs = (  
                "encoder_hidden_states",  
                "attention_mask",   
                "image_rotary_emb",  
            )  
            cross_attention = add_k_proj is not None
        elif module.is_cross_attention:
            q_proj, k_proj, v_proj = module.to_q, None, None
            add_q_proj, add_k_proj, add_v_proj, add_o_proj = None, module.to_k, module.to_v, None
            q_proj_rname, k_proj_rname, v_proj_rname = "to_q", "", ""
            add_q_proj_rname, add_k_proj_rname, add_v_proj_rname, add_o_proj_rname = "", "to_k", "to_v", ""
        else:
            q_proj, k_proj, v_proj = module.to_q, module.to_k, module.to_v
            add_q_proj = getattr(module, "add_q_proj", None)
            add_k_proj = getattr(module, "add_k_proj", None)
            add_v_proj = getattr(module, "add_v_proj", None)
            add_o_proj = getattr(module, "to_add_out", None)
            q_proj_rname, k_proj_rname, v_proj_rname = "to_q", "to_k", "to_v"
            add_q_proj_rname, add_k_proj_rname, add_v_proj_rname = "add_q_proj", "add_k_proj", "add_v_proj"
            add_o_proj_rname = "to_add_out"
        if getattr(module, "to_out", None) is not None:
            o_proj = module.to_out[0]
            o_proj_rname = "to_out.0"
            assert isinstance(o_proj, nn.Linear)
        elif parent is not None:
            assert isinstance(parent.module, FluxSingleTransformerBlock)
            assert isinstance(parent.module.proj_out, ConcatLinear)
            assert len(parent.module.proj_out.linears) == 2
            o_proj = parent.module.proj_out.linears[0]
            o_proj_rname = ".proj_out.linears.0"
        else:
            raise RuntimeError("Cannot find the output projection.")
        if isinstance(module.processor, DiffusionAttentionProcessor):
            with_rope = module.processor.rope is not None
        elif module.processor.__class__.__name__.startswith("Flux"):
            with_rope = True
        else:
            with_rope = False  # TODO: fix for other processors
        config = AttentionConfigStruct(
            hidden_size=q_proj.weight.shape[1],
            add_hidden_size=add_k_proj.weight.shape[1] if add_k_proj is not None else 0,
            inner_size=q_proj.weight.shape[0],
            num_query_heads=module.heads,
            num_key_value_heads=module.to_k.weight.shape[0] // (module.to_q.weight.shape[0] // module.heads),
            with_qk_norm=module.norm_q is not None,
            with_rope=with_rope,
            linear_attn=isinstance(module.processor, SanaLinearAttnProcessor2_0),
        )
        return DiffusionAttentionStruct(
            module=module,
            parent=parent,
            fname=fname,
            idx=idx,
            rname=rname,
            rkey=rkey,
            config=config,
            q_proj=q_proj,
            k_proj=k_proj,
            v_proj=v_proj,
            o_proj=o_proj,
            add_q_proj=add_q_proj,
            add_k_proj=add_k_proj,
            add_v_proj=add_v_proj,
            add_o_proj=add_o_proj,
            q=None,  # TODO: add q, k, v
            k=None,
            v=None,
            q_proj_rname=q_proj_rname,
            k_proj_rname=k_proj_rname,
            v_proj_rname=v_proj_rname,
            o_proj_rname=o_proj_rname,
            add_q_proj_rname=add_q_proj_rname,
            add_k_proj_rname=add_k_proj_rname,
            add_v_proj_rname=add_v_proj_rname,
            add_o_proj_rname=add_o_proj_rname,
            q_rname="",
            k_rname="",
            v_rname="",
        )
```

3) ValueError: Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined.

Potential Fix: app.diffusion.dataset.collect.calib.py

```python
def collect(config: DiffusionPtqRunConfig, dataset: datasets.Dataset):
    samples_dirpath = os.path.join(config.output.root, "samples")
    caches_dirpath = os.path.join(config.output.root, "caches")
    os.makedirs(samples_dirpath, exist_ok=True)
    os.makedirs(caches_dirpath, exist_ok=True)
    caches = []

    pipeline = config.pipeline.build()
    model = pipeline.unet if hasattr(pipeline, "unet") else pipeline.transformer
    assert isinstance(model, nn.Module)
    model.register_forward_hook(CollectHook(caches=caches), with_kwargs=True)

    batch_size = config.eval.batch_size
    print(f"In total {len(dataset)} samples")
    print(f"Evaluating with batch size {batch_size}")
    pipeline.set_progress_bar_config(desc="Sampling", leave=False, dynamic_ncols=True, position=1)
    for batch in tqdm(
        dataset.iter(batch_size=batch_size, drop_last_batch=False),
        desc="Data",
        leave=False,
        dynamic_ncols=True,
        total=(len(dataset) + batch_size - 1) // batch_size,
    ):
        filenames = batch["filename"]
        prompts = batch["prompt"]
        seeds = [hash_str_to_int(name) for name in filenames]
        generators = [torch.Generator(device=pipeline.device).manual_seed(seed) for seed in seeds]
        pipeline_kwargs = config.eval.get_pipeline_kwargs()

        task = config.pipeline.task
        control_root = config.eval.control_root
        if task in ["canny-to-image", "depth-to-image", "inpainting"]:
            controls = get_control(
                task,
                batch["image"],
                names=batch["filename"],
                data_root=os.path.join(
                    control_root, collect_config.dataset_name, f"{dataset.config_name}-{config.eval.num_samples}"
                ),
            )
            if task == "inpainting":
                pipeline_kwargs["image"] = controls[0]
                pipeline_kwargs["mask_image"] = controls[1]
            else:
                pipeline_kwargs["control_image"] = controls

        # Handle meta tensors by moving individual components  
        try:  
            pipeline = pipeline.to("cuda")  
        except NotImplementedError:  
            # Move individual pipeline components that have to_empty method  
            if hasattr(pipeline, 'transformer') and pipeline.transformer is not None:  
                try:  
                    pipeline.transformer = pipeline.transformer.to("cuda")  
                except NotImplementedError:  
                    pipeline.transformer = pipeline.transformer.to_empty(device="cuda")  

            if hasattr(pipeline, 'text_encoder') and pipeline.text_encoder is not None:  
                try:  
                    pipeline.text_encoder = pipeline.text_encoder.to("cuda")  
                except NotImplementedError:  
                    pipeline.text_encoder = pipeline.text_encoder.to_empty(device="cuda")  

            if hasattr(pipeline, 'text_encoder_2') and pipeline.text_encoder_2 is not None:  
                try:  
                    pipeline.text_encoder_2 = pipeline.text_encoder_2.to("cuda")  
                except NotImplementedError:  
                    pipeline.text_encoder_2 = pipeline.text_encoder_2.to_empty(device="cuda")  

            if hasattr(pipeline, 'vae') and pipeline.vae is not None:  
                try:  
                    pipeline.vae = pipeline.vae.to("cuda")  
                except NotImplementedError:  
                    pipeline.vae = pipeline.vae.to_empty(device="cuda")

        result_images = pipeline(prompt=prompts, generator=generators, **pipeline_kwargs).images
        num_guidances = (len(caches) // batch_size) // config.eval.num_steps
        num_steps = len(caches) // (batch_size * num_guidances)
        assert (
            len(caches) == batch_size * num_steps * num_guidances
        ), f"Unexpected number of caches: {len(caches)} != {batch_size} * {config.eval.num_steps} * {num_guidances}"
        for j, (filename, image) in enumerate(zip(filenames, result_images, strict=True)):
            image.save(os.path.join(samples_dirpath, f"{filename}.png"))
            for s in range(num_steps):
                for g in range(num_guidances):
                    c = caches[s * batch_size * num_guidances + g * batch_size + j]
                    c["filename"] = filename
                    c["step"] = s
                    c["guidance"] = g
                    c = tree_map(lambda x: process(x), c)
                    torch.save(c, os.path.join(caches_dirpath, f"{filename}-{s:05d}-{g}.pt"))
        caches.clear()
```

References

https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux.py#L266

https://github.com/nunchaku-tech/deepcompressor/blob/main/deepcompressor/nn/struct/attn.py

https://github.com/nunchaku-tech/nunchaku/blob/main/examples/flux.1-kontext-dev.py

https://github.com/nunchaku-tech/nunchaku/commit/b99fb8be615bc98c6915bbe06a1e0092cbc074a5

https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux/pipeline_flux_kontext.py

https://github.com/nunchaku-tech/deepcompressor/issues/91

---

# Dependencies
https://github.com/Dao-AILab/flash-attention

https://github.com/facebookresearch/xformers

https://github.com/openai/CLIP

https://github.com/THUDM/ImageReward

# Wheels

https://huggingface.co/datasets/siraxe/PrecompiledWheels_Torch-2.8-cu128-cp312

https://huggingface.co/lldacing/flash-attention-windows-wheel