[export] Update PT2 archive docs (#162308)

Summary: Minor updates based on the recent refactoring for weight saving and loading

Test Plan:
doc change only

Rollback Plan:

Differential Revision: D81821994

Pull Request resolved: https://github.com/pytorch/pytorch/pull/162308
Approved by: https://github.com/angelayi
This commit is contained in:
Yiming Zhou 2025-09-09 02:08:13 +00:00 committed by PyTorch MergeBot
parent 583bbf7761
commit a965f09793

View File

@ -22,18 +22,22 @@ The following is a sample archive. We will walk through the archive folder by fo
├── data ├── data
│ ├── aotinductor │ ├── aotinductor
│ │ └── model1 │ │ └── model1
│ │ ├── aotinductor_pickle_data.json │ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.kernel_metadata.json
│ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.cpp │ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.kernel.cpp
│ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.so │ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.wrapper_metadata.json
│ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.wrapper.cpp
│ │ ├── cf5ez6ifexr7i2hezzz4s7xfusj4wtisvu2gddeamh37bw6bghjw.wrapper.so
│ │ ├── cg7domx3woam3nnliwud7yvtcencqctxkvvcafuriladwxw4nfiv.cubin │ │ ├── cg7domx3woam3nnliwud7yvtcencqctxkvvcafuriladwxw4nfiv.cubin
│ │ └── cubaaxppb6xmuqdm4bej55h2pftbce3bjyyvljxbtdfuolmv45ex.cubin │ │ └── cubaaxppb6xmuqdm4bej55h2pftbce3bjyyvljxbtdfuolmv45ex.cubin
│ ├── weights │ ├── weights
│ │ ├── model1_model_param_config.json │ │ ├── model1_weights_config.json
│ │ ├── model2_weights_config.json
│ │ ├── weight_0 │ │ ├── weight_0
│ │ ├── weight_1 │ │ ├── weight_1
│ │ ├── weight_2 │ │ ├── weight_2
│ └── constants │ └── constants
│ │ ├── model1_model_constants_config.json │ │ ├── model1_constants_config.json
│ │ ├── model2_constants_config.json
│ │ ├── tensor_0 │ │ ├── tensor_0
│ │ ├── tensor_1 │ │ ├── tensor_1
│ │ ├── custom_obj_0 │ │ ├── custom_obj_0
@ -67,11 +71,12 @@ example, compilation artifacts for the `model1` model on A100 and H100 will be
saved in `model1-a100` and `model1-h100` folders separately. saved in `model1-a100` and `model1-h100` folders separately.
The folder typically contains The folder typically contains
* `<uuid>.so`: Dynamic library compiled from <uuid>.cpp. * `<uuid>.wrapper.so`: Dynamic library compiled from <uuid>.cpp.
* `<uuid>.cpp`: AOTInductor generated cpp wrapper file. * `<uuid>.wrapper.cpp`: AOTInductor generated cpp wrapper file.
* `<uuid>.kernel.cpp`: AOTInductor generated cpp kernel file.
* `*.cubin`: Triton kernels compiled from triton codegen kernels * `*.cubin`: Triton kernels compiled from triton codegen kernels
* `<uuid>.wrapper_metadata.json`: Metadata which was passed in from the `aot_inductor.metadata` inductor config
* (optional) `<uuid>.json`: External fallback nodes for custom ops to be executed by `ProxyExecutor`, serialized according to `ExternKernelNode` struct. If the model doesnt use custom ops/ProxyExecutor, this file would be omitted. * (optional) `<uuid>.json`: External fallback nodes for custom ops to be executed by `ProxyExecutor`, serialized according to `ExternKernelNode` struct. If the model doesnt use custom ops/ProxyExecutor, this file would be omitted.
* `<uuid>_metadata.json`: Metadata which was passed in from the `aot_inductor.metadata` inductor config
### Weights ### Weights
@ -79,16 +84,16 @@ Path: `/data/weights/*`
Model parameters and buffers are saved in the `/data/weights/` folder. Each Model parameters and buffers are saved in the `/data/weights/` folder. Each
tensor is saved as a separated file. The file only contains the raw data blob, tensor is saved as a separated file. The file only contains the raw data blob,
tensor metadata are saved separately in the tensor metadata and mapping from model weight FQN to saved raw data blob are saved separately in the
`<model_name>_model_param_config.json`. `<model_name>_weights_config.json`.
### Constants ### Constants
Path: `/data/constants/*` Path: `/data/constants/*`
TensorConstants, non-persistent buffers and TorchBind objects are saved in the TensorConstants, non-persistent buffers and TorchBind objects are saved in the
`/data/constants/` folder. Metadata is saved separately in the `/data/constants/` folder. Metadata and mapping from model constant FQN to saved raw data blob are saved separately in the
`<model_name>_model_constants_config.json` `<model_name>_constants_config.json`
### Sample Inputs ### Sample Inputs