Update dependencies and fix convert_model (#458)

* feat: Update dependencies in Cargo.toml

Update the dependencies in Cargo.toml to their latest versions:
- rust_tokenizers: 8.1.1
- tch: 0.16.0 (with features = ["download-libtorch"])
- serde_json: 1
- serde: 1 (with features = ["derive"])
- ordered-float: 4.2.0
- uuid: 1 (with features = ["v4"])
- thiserror: 1
- half: 2
- regex: 1.6
- cached-path: 0.6 (with default-features = false and optional = true)
- dirs: 5 (optional = true)
- lazy_static: 1 (optional = true)
- ort: 1.16.3 (optional = true, default-features = false, features = ["half"])
- ndarray: 0.15 (optional = true)
- tokenizers: 0.19.1 (optional = true, default-features = false, features = ["onig"])

```

* chore: Update .gitignore and requirements.txt, and improve convert_model.py

Update .gitignore to exclude the /models/ and /.venv/ directories, and the convert_model.log file.

Remove the requirements.txt file.

In convert_model.py:
- Add a new function, `zipfile_factory`, to handle zip file creation.
- Update the logger configuration to log debug messages to a file named `convert_model.log`.

* delete duplicate requirements file

* update CI req file path

* missing requests dependency

---------

Co-authored-by: Abdulrhman Alkhodiry <[email protected]>
Co-authored-by: Guillaume Becquin <[email protected]>

.github/workflows/continuous-integration.yml

@@ -174,7 +174,7 @@ jobs:
         with:
           python-version: '3.10'
       - run: |
-          pip install -r requirements.txt --progress-bar off
+          pip install -r ./utils/requirements.txt --progress-bar off
           python ./utils/download-dependencies_distilbert.py
 
   fmt:

.gitignore

@@ -17,4 +17,6 @@ Cargo.lock
 
 /target
 #**/*.rs.bk
-/resources/
+/models/
+/.venv/
+convert_model.log

Cargo.toml

@@ -76,10 +76,10 @@ features = ["doc-only"]
 
 [dependencies]
 rust_tokenizers = "8.1.1"
-tch = "0.15.0"
+tch = { version = "0.16.0", features = ["download-libtorch"] }
 serde_json = "1"
 serde = { version = "1", features = ["derive"] }
-ordered-float = "3"
+ordered-float = "4.2.0"
 uuid = { version = "1", features = ["v4"] }
 thiserror = "1"
 half = "2"
@@ -88,19 +88,26 @@ regex = "1.6"
 cached-path = { version = "0.6", default-features = false, optional = true }
 dirs = { version = "5", optional = true }
 lazy_static = { version = "1", optional = true }
-ort = {version="~1.15.2", optional = true, default-features = false, features = ["half"]}
-ndarray = {version="0.15", optional = true}
-tokenizers = {version="0.15", optional=true, default-features = false, features = ["onig"]}
+ort = { version = "1.16.3", optional = true, default-features = false, features = [
+    "half",
+] }
+ndarray = { version = "0.15", optional = true }
+tokenizers = { version = "0.19.1", optional = true, default-features = false, features = [
+    "onig",
+] }
 
 [dev-dependencies]
 anyhow = "1"
 csv = "1"
 criterion = "0.5"
 tokio = { version = "1.35", features = ["sync", "rt-multi-thread", "macros"] }
 tempfile = "3"
-itertools = "0.12"
-tracing-subscriber = { version = "0.3", default-features = false, features = [ "env-filter", "fmt" ] }
-ort = {version="~1.15.5", features = ["load-dynamic"]}
+itertools = "0.13.0"
+tracing-subscriber = { version = "0.3", default-features = false, features = [
+    "env-filter",
+    "fmt",
+] }
+ort = { version = "1.16.3", features = ["load-dynamic"] }
 
 [[example]]
 name = "onnx-masked-lm"

src/models/bart/decoder.rs

@@ -340,6 +340,7 @@ impl BartDecoder {
     }
 }
 
+#[allow(dead_code)]
 ///Container holding a BART decoder output
 pub struct BartDecoderOutput {
     /// last decoder layer hidden state

src/models/prophetnet/encoder.rs

@@ -210,6 +210,7 @@ impl ProphetNetEncoder {
     }
 }
 
+#[allow(dead_code)]
 /// Container for the ProphetNet encoder output.
 pub struct ProphetNetEncoderOutput {
     /// Last hidden states from the model

src/models/reformer/encoder.rs

@@ -143,6 +143,7 @@ impl ChunkReformerFeedForward {
     }
 }
 
+#[allow(dead_code)]
 pub struct ReformerLayerOutput {
     pub attention_output: Tensor,
     pub hidden_states: Tensor,

src/models/reformer/reformer_model.rs

@@ -207,6 +207,7 @@ impl ReformerLMHead {
     }
 }
 
+#[allow(dead_code)]
 pub struct PaddedReformerInput {
     pub input_ids: Option<Tensor>,
     pub input_embeds: Option<Tensor>,

src/models/t5/encoder.rs

@@ -541,6 +541,7 @@ impl T5Stack {
     }
 }
 
+#[allow(dead_code)]
 pub struct T5BlockOutput {
     pub hidden_states: Tensor,
     pub self_attention_weights: Option<Tensor>,

utils/convert_model.py

@@ -50,13 +50,20 @@
 import zipfile
 from pathlib import Path
 from typing import Dict
-
+import os
 import numpy as np
 import torch
 from numpy.lib.format import write_array
-from numpy.lib.npyio import zipfile_factory
+# from numpy.lib.npyio import zipfile_factory
 from torch import Tensor
 
+def zipfile_factory(file, *args, **kwargs):
+    if not hasattr(file, 'read'):
+        file = os.fspath(file)
+    import zipfile
+    kwargs['allowZip64'] = True
+    kwargs['compresslevel'] = 4
+    return zipfile.ZipFile(file, *args, **kwargs)
 
 def get_bf16_repr(input_tensor: torch.Tensor) -> np.ndarray:
     """Convert a bfloat16 tensor to an equivalent byte representation in Numpy.
@@ -125,6 +132,12 @@ def append_to_zipf(
         help="Use this flag to enable automatic download of the libtorch library.",
     )
     args = parser.parse_args()
+
+    logger = logging.getLogger('convert_model')
+    logger.setLevel(logging.DEBUG)
+    fh = logging.FileHandler('convert_model.log')
+    fh.setLevel(logging.DEBUG)
+    logger.addHandler(fh)
 
     target_folder = Path(args.source_file[0]).parent
     with zipfile_factory(
@@ -133,7 +146,7 @@ def append_to_zipf(
         for source_file_or_pattern in args.source_file:
             source_files = glob.glob(source_file_or_pattern)
             for source_file in source_files:
-                logging.info(f"Processing source file {source_file}...")
+                logger.info(f"Processing source file {source_file}")
                 nps = {}
                 source_file = Path(source_file)
                 weights = torch.load(str(source_file), map_location="cpu")
@@ -168,11 +181,11 @@ def append_to_zipf(
                             )
                         else:
                             nps[k] = np.ascontiguousarray(tensor)
-                        logging.info(
+                        logger.info(
                             f"converted {k} - {str(sys.getsizeof(nps[k]))} bytes"
                         )
                     else:
-                        logging.info(f"skipped non-tensor object: {k}")
+                        logger.info(f"skipped non-tensor object: {k}")
             append_to_zipf(nps, output_zipfile)
 
     source = str(target_folder / "model.npz")

utils/requirements.txt

@@ -0,0 +1,11 @@
+filelock==3.15.3
+fsspec==2024.6.0
+Jinja2==3.1.4
+MarkupSafe==2.1.5
+mpmath==1.3.0
+networkx==3.3
+numpy==2.0.0
+sympy==1.12.1
+torch==2.3.1
+typing_extensions==4.12.2
+requests==2.32.0