File size: 18,105 Bytes
a9bd396 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 | <!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Model debugging toolboxes
This page lists all the debugging and model adding tools used by the library, as well as the utility functions it
provides for it.
Most of those are only useful if you are adding new models in the library.
## Model addition debuggers
### Model addition debugger - context manager for model adders
This context manager is a power user tool intended for model adders. It tracks all forward calls within a model forward
and logs a slice of each input and output on a nested JSON. To note, this context manager enforces `torch.no_grad()`.
### Rationale
When porting models to transformers, even from python to python, model adders often have to do a lot of manual
operations, involving saving and loading tensors, comparing dtypes, etc. This small tool can hopefully shave off some
time.
### Usage
Add this context manager as follows to debug a model:
```python
import torch
from PIL import Image
import requests
from transformers import LlavaProcessor, LlavaForConditionalGeneration
from transformers.model_debugging_utils import model_addition_debugger_context
torch.random.manual_seed(673)
# load pretrained model and processor
model_id = "llava-hf/llava-1.5-7b-hf"
processor = LlavaProcessor.from_pretrained(model_id)
model = LlavaForConditionalGeneration.from_pretrained(model_id)
# create random image input
random_image = Image.fromarray(torch.randint(0, 256, (224, 224, 3), dtype=torch.uint8).numpy())
# prompt
prompt = "<image>Describe this image."
# process inputs
inputs = processor(text=prompt, images=random_image, return_tensors="pt")
# call forward method (not .generate!)
with model_addition_debugger_context(
model,
debug_path="optional_path_to_your_directory",
do_prune_layers=False # This will output ALL the layers of a model.
):
output = model.forward(**inputs)
```
### Reading results
The debugger generates two files from the forward call, both with the same base name, but ending either with
`_SUMMARY.json` or with `_FULL_TENSORS.json`.
The first one will contain a summary of each module's _input_ and _output_ tensor values and shapes.
```json
{
"module_path": "MolmoForConditionalGeneration",
"inputs": {
"args": [],
"kwargs": {
"input_ids": {
"shape": "torch.Size([1, 589])",
"dtype": "torch.int64"
},
"attention_mask": {
"shape": "torch.Size([1, 589])",
"dtype": "torch.int64"
},
"pixel_values": {
"shape": "torch.Size([1, 5, 576, 588])",
"dtype": "torch.float32",
"mean": "tensor(-8.9514e-01, device='cuda:0')",
"std": "tensor(9.2586e-01, device='cuda:0')",
"min": "tensor(-1.7923e+00, device='cuda:0')",
"max": "tensor(1.8899e+00, device='cuda:0')"
}
},
"children": [
{
"module_path": "MolmoForConditionalGeneration.language_model.model.embed_tokens",
"inputs": {
"args": [
{
"shape": "torch.Size([1, 589])",
"dtype": "torch.int64"
}
]
},
"outputs": {
"shape": "torch.Size([1, 589, 3584])",
"dtype": "torch.float32",
"mean": "tensor(6.5460e-06, device='cuda:0')",
"std": "tensor(2.3807e-02, device='cuda:0')",
"min": "tensor(-3.3398e-01, device='cuda:0')",
"max": "tensor(3.9453e-01, device='cuda:0')"
}
},
{
"module_path": "MolmoForConditionalGeneration.vision_tower",
"inputs": {
"args": [
{
"shape": "torch.Size([5, 1, 576, 588])",
"dtype": "torch.float32",
"mean": "tensor(-8.9514e-01, device='cuda:0')",
"std": "tensor(9.2586e-01, device='cuda:0')",
"min": "tensor(-1.7923e+00, device='cuda:0')",
"max": "tensor(1.8899e+00, device='cuda:0')"
}
],
"kwargs": {
"output_hidden_states": "True"
}
},
"children": [
{ ... and so on
```
The `_FULL_TENSORS.json` file will display a full view of all tensors, which is useful for comparing two files.
```json
"pixel_values": {
"shape": "torch.Size([1, 5, 576, 588])",
"dtype": "torch.float32",
"value": [
"tensor([[[[-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" ...,",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00]],",
"",
" [[-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" ...,",
" [-1.4857e+00, -1.4820e+00, -1.2100e+00, ..., -6.0979e-01, -5.9650e-01, -3.8527e-01],",
" [-1.6755e+00, -1.7221e+00, -1.4518e+00, ..., -7.5577e-01, -7.4658e-01, -5.5592e-01],",
" [-7.9957e-01, -8.2162e-01, -5.7014e-01, ..., -1.3689e+00, -1.3169e+00, -1.0678e+00]],",
"",
" [[-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" ...,",
" [-3.0322e-01, -5.0645e-01, -5.8436e-01, ..., -6.2439e-01, -7.9160e-01, -8.1188e-01],",
" [-4.4921e-01, -6.5653e-01, -7.2656e-01, ..., -3.4702e-01, -5.2146e-01, -5.1326e-01],",
" [-3.4702e-01, -5.3647e-01, -5.4170e-01, ..., -1.0915e+00, -1.1968e+00, -1.0252e+00]],",
"",
" [[-1.1207e+00, -1.2718e+00, -1.0678e+00, ..., 1.2013e-01, -1.3126e-01, -1.7197e-01],",
" [-6.9738e-01, -9.1166e-01, -8.5454e-01, ..., -5.5050e-02, -2.8134e-01, -4.2793e-01],",
" [-3.4702e-01, -5.5148e-01, -5.8436e-01, ..., 1.9312e-01, -8.6235e-02, -2.1463e-01],",
" ...,",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00]],",
"",
" [[-1.0039e+00, -9.5669e-01, -6.5546e-01, ..., -1.4711e+00, -1.4219e+00, -1.1389e+00],",
" [-1.0039e+00, -9.5669e-01, -6.5546e-01, ..., -1.7193e+00, -1.6771e+00, -1.4091e+00],",
" [-1.6317e+00, -1.6020e+00, -1.2669e+00, ..., -1.2667e+00, -1.2268e+00, -8.9720e-01],",
" ...,",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
" [-1.7923e+00, -1.7521e+00, -1.4802e+00, ..., -1.7923e+00, -1.7521e+00, -1.4802e+00]]]], device='cuda:0')"
],
"mean": "tensor(-8.9514e-01, device='cuda:0')",
"std": "tensor(9.2586e-01, device='cuda:0')",
"min": "tensor(-1.7923e+00, device='cuda:0')",
"max": "tensor(1.8899e+00, device='cuda:0')"
},
```
#### Saving tensors to disk
Some model adders may benefit from logging full tensor values to disk to support, for example, numerical analysis
across implementations.
Set `use_repr=False` to write tensors to disk using [SafeTensors](https://huggingface.co/docs/safetensors/en/index).
```python
with model_addition_debugger_context(
model,
debug_path="optional_path_to_your_directory",
do_prune_layers=False,
use_repr=False, # Defaults to True
):
output = model.forward(**inputs)
```
When using `use_repr=False`, tensors are written to the same disk location as the `_SUMMARY.json` and
`_FULL_TENSORS.json` files. The `value` property of entries in the `_FULL_TENSORS.json` file will contain a relative
path reference to the associated `.safetensors` file. Each tensor is written to its own file as the `data` property of
the state dictionary. File names are constructed using the `module_path` as a prefix with a few possible postfixes that
are built recursively.
* Module inputs are denoted with the `_inputs` and outputs by `_outputs`.
* `list` and `tuple` instances, such as `args` or function return values, will be postfixed with `_{index}`.
* `dict` instances will be postfixed with `_{key}`.
### Comparing between implementations
Once the forward passes of two models have been traced by the debugger, one can compare the `json` output files. See
below: we can see slight differences between these two implementations' key projection layer. Inputs are mostly
identical, but not quite. Looking through the file differences makes it easier to pinpoint which layer is wrong.
### Limitations and scope
This feature will only work for torch-based models. Models relying heavily on external kernel calls may work, but trace will
probably miss some things. Regardless, any python implementation that aims at mimicking another implementation can be
traced once instead of reran N times with breakpoints.
If you pass `do_prune_layers=False` to your model debugger, ALL the layers will be outputted to `json`. Else, only the
first and last layer will be shown. This is useful when some layers (typically cross-attention) appear only after N
layers.
[[autodoc]] model_addition_debugger_context
## Analyzer of skipped tests
### Scan skipped tests - for model adders and maintainers
This small util is a power user tool intended for model adders and maintainers. It lists all test methods
existing in `test_modeling_common.py`, inherited by all model tester classes, and scans the repository to measure
how many tests are being skipped and for which models.
### Rationale
When porting models to transformers, tests fail as they should, and sometimes `test_modeling_common` feels irreconcilable with the peculiarities of our brand new model. But how can we be sure we're not breaking everything by adding a seemingly innocent skip?
This utility:
- scans all test_modeling_common methods
- looks for times where a method is skipped
- returns a summary json you can load as a DataFrame/inspect
**For instance test_inputs_embeds is skipped in a whooping 39% proportion at the time of writing this util.**
### Usage
You can run the skipped test analyzer in two ways:
#### Full scan (default)
From the root of `transformers` repo, scans all common test methods and outputs the results to a JSON file (default: `all_tests_scan_result.json`).
```bash
python utils/scan_skipped_tests.py --output_dir path/to/output
```
- `--output_dir` (optional): Directory where the JSON results will be saved. Defaults to the current directory.
**Example output:**
```text
🔬 Parsing 331 model test files once each...
📝 Aggregating 224 tests...
(224/224) test_update_candidate_strategy_with_matches_1es_3d_is_nonecodet_schedule_fa_kwargs
✅ Scan complete.
📄 JSON saved to /home/pablo/git/transformers/all_tests_scan_result.json
```
And it will generate `all_tests_scan_result.json` file that you can inspect. The JSON is indexed by method name, and each entry follows this schema, indicating the origin as well (from `common`or `GenerationMixin`.)
```json
{
"<method_name>": {
"origin": "<test suite>"
"models_ran": ["<model_name>", ...],
"models_skipped": ["<model_name>", ...],
"skipped_proportion": <float>,
"reasons_skipped": ["<model_name>: <reason>",
...
]
},
...
}
```
Which you can visualise as above with e.g. `pandas`
```python
df = pd.read_json('all_tests_scan_result.json').T
df.sort_values(by=['skipped_proportion'], ascending=False)
```
### Scan a single test method
You can focus on a specific test method using `--test_method_name`:
```bash
python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds --output_dir path/to/output
```
- `--test_method_name`: Name of the test method to scan (e.g., `test_inputs_embeds`).
- `--output_dir` (optional): Directory where the JSON result will be saved.
**Example output:**
```bash
$ python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds
🔬 Parsing 331 model test files once each...
== test_inputs_embeds ==
Ran : 199/323
Skipped : 124/323 (38.4%)
- aimv2: Aimv2 does not use inputs_embeds
- align: Inputs_embeds is tested in individual model tests
- altclip: Inputs_embeds is tested in individual model tests
- audio_spectrogram_transformer: AST does not use inputs_embeds
- beit: BEiT does not use inputs_embeds
- bit: Bit does not use inputs_embeds
- blip: Blip does not use inputs_embeds
- blip_2: Inputs_embeds is tested in individual model tests
- bridgetower:
- canine: CANINE does not have a get_input_embeddings() method.
- ...
📄 JSON saved to /home/pablo/git/transformers/scan_test_inputs_embeds.json
```
## Modular model detector
### Code similarity analyzer - for model adders
This utility analyzes code similarities between model implementations to identify opportunities for modularization. It compares a new or existing modeling file against all models in the library using embedding-based and token-based similarity metrics.
### Rationale
When adding a new model to transformers, many components (attention layers, MLPs, outputs, etc.) may already exist in similar form in other models. Instead of implementing everything from scratch, model adders can identify which existing classes are similar and potentially reusable through modularization.
The tool computes two similarity scores:
- **Embedding score**: Uses semantic code embeddings (via `Qwen/Qwen3-Embedding-4B`) to detect functionally similar code even with different naming
- **Jaccard score**: Measures token set overlap to identify structurally similar code patterns
A score of 1.00 means the code is identical.
### Usage
From the root of the `transformers` repository:
```bash
python utils/modular_model_detector.py --modeling-file path/to/modeling_file.py
```
The tool will automatically download the pre-built index from the Hub (requires RAM/VRAM for the embedding model).
**Example output:**
```text
Loading checkpoint shards: 100%|████████████████████| 2/2 [00:00<00:00, 33.62it/s]
encoding 21 query definitions with Qwen/Qwen3-Embedding-4B (device=cuda, batch=16, max_length=4096)
stuff.py::Beit3ImageTextMatchingOutput:
embedding:
blip_2::Blip2ImageTextMatchingModelOutput (0.9994)
chinese_clip::ChineseCLIPOutput (0.9818)
owlvit::OwlViTOutput (0.9818)
jaccard:
owlv2::Owlv2Output (0.9667)
metaclip_2::MetaClip2Output (0.9667)
altclip::AltCLIPOutput (0.9667)
intersection:
blip::BlipOutput
owlvit::OwlViTOutput
stuff.py::Beit3MLP:
embedding:
efficientloftr::EfficientLoFTRMLP (0.9718)
seggpt::SegGptMlp (0.9650)
jaccard:
chinese_clip::ChineseCLIPTextSelfOutput (0.5294)
bert::BertSelfOutput (0.5294)
intersection:
```
The `intersection` field shows classes that appear in both top-5 results, indicating high confidence for modularization candidates.
### Building a custom index
To rebuild the index from your local codebase (useful after adding new models or using a different embedding model):
```bash
python utils/modular_model_detector.py --build
```
To push the rebuilt index to a Hub dataset:
```bash
python utils/modular_model_detector.py --build --push-new-index --hub-dataset your-org/your-dataset
```
### Options
- `--modeling-file`: Path to the modeling file to analyze
- `--build`: Build the code similarity index from all modeling files in `src/transformers/models/`
- `--push-new-index`: After building, push the index to a Hub dataset (requires `--build`)
- `--hub-dataset`: Hub dataset repository ID to pull/push the index (default: `hf-internal-testing/transformers_code_embeddings`)
### Limitations
This tool requires GPU/CPU resources to run the embedding model (`Qwen/Qwen3-Embedding-4B`). The pre-built index is downloaded from the Hub by default, which requires an internet connection on first use.
Results are suggestions based on code similarity and should be manually reviewed before modularization. High similarity scores don't guarantee perfect compatibility.
|