Skip to content

TimesFM #2135

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
kashif wants to merge 13 commits into from
Closed

TimesFM #2135

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
4b82c3c
initial
kashif Dec 22, 2024
f5b035f
merge main in branch
echarlaix Feb 28, 2025
f5509ff
fix style
echarlaix Feb 28, 2025
5c2fee5
add fixes
echarlaix Feb 28, 2025
c373ae0
remove time series normalized config
echarlaix Feb 28, 2025
9329cb0
update normalized config
echarlaix Feb 28, 2025
a41da57
Merge branch 'main' into timesfm
echarlaix Apr 24, 2025
5a50917
remove mapping context len
echarlaix Apr 24, 2025
af6a2d3
fix export
echarlaix Apr 24, 2025
0ece0c5
add test
echarlaix Apr 24, 2025
891512a
move dummy input generator to utils
echarlaix Apr 24, 2025
6d21600
upgrade min transformers version
echarlaix Apr 24, 2025
d75e046
Merge branch 'main' into timesfm
echarlaix Apr 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions optimum/exporters/onnx/model_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,7 @@
NormalizedTimeSeriesForecastingConfig,
NormalizedVisionConfig,
PerceiverDummyInputGenerator,
TimesFMDummyInputGenerator,
VitPoseDummyInputGenerator,
is_diffusers_available,
is_diffusers_version,
Expand Down Expand Up @@ -2641,6 +2642,25 @@ class EncoderDecoderOnnxConfig(EncoderDecoderBaseOnnxConfig):
DEFAULT_ONNX_OPSET = 14 # uses SDPA in Transformers, hence opset>=14.


class TimesFMOnnxConfig(OnnxConfig):
NORMALIZED_CONFIG_CLASS = NormalizedTimeSeriesForecastingConfig
MIN_TRANSFORMERS_VERSION = version.parse("4.52.0")
DUMMY_INPUT_GENERATOR_CLASSES = (TimesFMDummyInputGenerator,)
DEFAULT_ONNX_OPSET = 14 # uses SDPA in Transformers, needs opset>=14

@property
def inputs(self) -> Dict[str, Dict[int, str]]:
return {"past_values": {0: "batch_size", 1: "sequence_length"}}

@echarlaix echarlaix Apr 25, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

an issue is remaining before we can merge : it seems that the resulting onnx graph doesn't considered the batch_size as dynamic (won't accept one different than the one given during export) likely coming from :

https://github.com/huggingface/transformers/blob/79d4bc761d9da7a125b65a3b8c09441608d05436/src/transformers/models/timesfm/modeling_timesfm.py#L791

@kashif any chance we can add the modification directly in the transformers modeling ? otherwise we need to patch the model before export

@kashif kashif Apr 25, 2025

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

checking but i believe so

@xenova xenova Apr 25, 2025

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Indeed, that does look to be the source of the loss of dynamic shapes. torch.onnx.export doesn't play well with python list operations, so we'd need to refactor it to use tensor operations, if possible.


@property
def outputs(self) -> Dict[str, Dict[int, str]]:
return {
"last_hidden_state": {0: "batch_size"},
"mean_predictions": {0: "batch_size"},
"full_predictions": {0: "batch_size"},
}


class PatchTSTOnnxConfig(OnnxConfig):
NORMALIZED_CONFIG_CLASS = NormalizedTimeSeriesForecastingConfig
DUMMY_INPUT_GENERATOR_CLASSES = (DummyPatchTSTInputGenerator,)
Expand Down
5 changes: 5 additions & 0 deletions optimum/exporters/tasks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -329,6 +329,7 @@ class TasksManager:
("pt", "visual-bert", "question-answering"): ("transformers", "VisualBertForQuestionAnswering"),
# VisionEncoderDecoderModel is not registered in AutoModelForDocumentQuestionAnswering
("pt", "vision-encoder-decoder", "document-question-answering"): ("transformers", "VisionEncoderDecoderModel"),
("pt", "timesfm", "time-series-forecasting"): ("transformers", "TimesFmModelForPrediction"),
("pt", "vitpose", "keypoint-detection"): ("transformers", "VitPoseForPoseEstimation"),
}

Expand Down Expand Up @@ -996,6 +997,10 @@ class TasksManager:
"text-classification",
onnx="Qwen2OnnxConfig",
),
"timesfm": supported_tasks_mapping(
"time-series-forecasting",
onnx="TimesFMOnnxConfig",
),
"llama": supported_tasks_mapping(
"feature-extraction",
"feature-extraction-with-past",
Expand Down
1 change: 1 addition & 0 deletions optimum/utils/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,7 @@
MistralDummyPastKeyValuesGenerator,
MultiQueryPastKeyValuesGenerator,
PerceiverDummyInputGenerator,
TimesFMDummyInputGenerator,
VitPoseDummyInputGenerator,
)
from .modeling_utils import recurse_getattr, recurse_setattr
Expand Down
25 changes: 25 additions & 0 deletions optimum/utils/input_generators.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1658,3 +1658,28 @@ class PerceiverDummyInputGenerator(DummyVisionStaticInputGenerator):

class VitPoseDummyInputGenerator(DummyVisionStaticInputGenerator):
pass


class TimesFMDummyInputGenerator(DummyInputGenerator):
SUPPORTED_INPUT_NAMES = ("past_values",)

def __init__(
self,
task: str,
normalized_config: NormalizedConfig,
batch_size: int = DEFAULT_DUMMY_SHAPES["batch_size"],
**kwargs,
):
self.task = task
self.normalized_config = normalized_config
self.batch_size = batch_size
self.context_length = normalized_config.context_length

def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int64", float_dtype: str = "fp32"):
return self.random_float_tensor(
shape=[self.batch_size, self.context_length],
min_value=-1,
max_value=1,
framework=framework,
dtype=float_dtype,
)
1 change: 1 addition & 0 deletions tests/exporters/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,6 +171,7 @@
"swin2sr": "hf-internal-testing/tiny-random-Swin2SRModel",
"t5": "hf-internal-testing/tiny-random-t5",
"table-transformer": "hf-internal-testing/tiny-random-TableTransformerModel",
"timesfm": "huggingface/timesfm-tourism-monthly",
"vit": "hf-internal-testing/tiny-random-vit",
"vit-mae": "hf-internal-testing/tiny-random-ViTMAEModel",
"vit-msn": "hf-internal-testing/tiny-random-ViTMSNForImageClassification",
Expand Down