fix ci

Author: xadupre

.github/workflows/ci.yml

@@ -17,7 +17,7 @@ jobs:
         os: [ubuntu-latest]
         python: ['3.11', '3.12']
         transformers: ['4.48', '4.50', 'main']
-        torch: ['main']
+        torch: ['2.6', 'main']
 
     steps:
       - uses: actions/checkout@v3

CHANGELOGS.rst

@@ -4,7 +4,9 @@ Change Logs
 0.2.1
 +++++
 
-* :pr:`16`: refactors patches
+* :pr:`16`: refactors patches, add model Phi2, implements
+  a tweak to raise an exception with a dynamic dimension
+  becomes static when exporting a model
 
 0.2.0
 +++++

_doc/examples/plot_export_tiny_llm.py

@@ -125,7 +125,7 @@ def _forward_(*args, _f=None, **kwargs):
 
 try:
     ep = torch.export.export(
-        untrained_model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes
+        untrained_model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes, strict=False
     )
     print("It worked:")
     print(ep)
@@ -159,7 +159,9 @@ def _forward_(*args, _f=None, **kwargs):
 # And Let's finally export.
 
 try:
-    ep = torch.export.export(model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes)
+    ep = torch.export.export(
+        model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes, strict=False
+    )
     print("It worked:")
     print(ep)
 except Exception as e:

_doc/examples/plot_export_tiny_llm_patched.py

@@ -29,7 +29,8 @@
       specified at `dynamic_shapes['past_key_values']`
       to non-tensor type <class 'transformers.cache_utils.DynamicCache'>
       at `inputs['past_key_values']` (expected None)
-  For more information about this error, see: https://pytorch.org/docs/main/generated/exportdb/index.html#dynamic-shapes-validation
+  For more information about this error,
+  see: https://pytorch.org/docs/main/generated/exportdb/index.html#dynamic-shapes-validation
 
 With ``transformers==4.50``, it shows the following:
 
@@ -67,8 +68,9 @@
 import torch
 import transformers
 from onnx_diagnostic import doc
+from onnx_diagnostic.cache_helpers import is_cache_dynamic_registered
 from onnx_diagnostic.helpers import string_type
-from onnx_diagnostic.torch_export_patches.onnx_export_errors import bypass_export_some_errors
+from onnx_diagnostic.torch_export_patches import bypass_export_some_errors
 from onnx_diagnostic.torch_models.llms import get_tiny_llm
 
 
@@ -92,14 +94,25 @@
 pprint.pprint(dynamic_shapes)
 
 # %%
-# We are ready to export.
+# Before exporting, we check :class:`transformers.cache_utils.DynamicCache`
+# can serialized and deserialized otherwise :func:`torch.export.export`
+# fails.
 
-with bypass_export_some_errors(patch_transformers=True) as modificator:
+print("-- DynamicCache registered: ", is_cache_dynamic_registered())
+
+# %%
+# If they are not registered, function
+# func:`onnx_diagnostic.torch_export_patches.bypass_export_some_errors`
+# should take care of it. Then we export.
+
+with bypass_export_some_errors(patch_transformers=True, verbose=10) as modificator:
+    assert is_cache_dynamic_registered()  # it must be true here
     ep = torch.export.export(
         untrained_model,
         (),
         kwargs=modificator(cloned_inputs),
         dynamic_shapes=dynamic_shapes,
+        strict=False,  # mandatory for torch==2.6
     )
     print("It worked:")
     print(ep)
@@ -114,12 +127,13 @@
 
 cloned_inputs = copy.deepcopy(inputs)
 
-with bypass_export_some_errors(patch_transformers=True) as modificator:
+with bypass_export_some_errors(patch_transformers=True, verbose=10) as modificator:
     ep = torch.export.export(
         model,
         (),
         kwargs=modificator(cloned_inputs),
         dynamic_shapes=dynamic_shapes,
+        strict=False,  # mandatory for torch==2.6
     )
     print("It worked:")
     print(ep)

_unittests/ut_torch_models/test_tiny_llms_onnx.py

@@ -6,6 +6,7 @@
     ExtTestCase,
     ignore_warnings,
     hide_stdout,
+    has_torch,
     requires_transformers,
 )
 from onnx_diagnostic.torch_models.llms import get_tiny_llm
@@ -35,6 +36,9 @@ def test_onnx_export_tiny_llm_official(self):
             dynamo=True,
             optimize=True,
         )
+        # There are some discrepancies with torch==2.6
+        if not has_torch("2.7"):
+            raise unittest.SkipTest("discrepancies observed with torch<2.7")
         self.assert_onnx_disc(
             inspect.currentframe().f_code.co_name, ep.model_proto, model, inputs, verbose=1
         )
@@ -96,6 +100,9 @@ def test_bypass_onnx_export_tiny_llm_official_full(self):
                 dynamo=True,
                 optimize=True,
             )
+        # There are some discrepancies with torch==2.6
+        if not has_torch("2.7"):
+            raise unittest.SkipTest("discrepancies observed with torch<2.7")
         self.assert_onnx_disc(
             inspect.currentframe().f_code.co_name, ep.model_proto, model, inputs, verbose=1
         )

onnx_diagnostic/cache_helpers.py

@@ -4,6 +4,28 @@
 import transformers
 import transformers.cache_utils
 
+
+def is_cache_dynamic_registered() -> bool:
+    """
+    Tells class :class:`transformers.cache_utils.DynamicCache` can be
+    serialized and deserialized. Only then, :func:`torch.export.export`
+    can export a model.
+    """
+    bsize, nheads, slen, dim = 2, 4, 3, 7
+    cache = make_dynamic_cache(
+        [
+            (
+                torch.randn(bsize, nheads, slen, dim),
+                torch.randn(bsize, nheads, slen, dim),
+            )
+            for i in range(2)
+        ]
+    )
+    values, spec = torch.utils._pytree.tree_flatten(cache)
+    cache2 = torch.utils._pytree.tree_unflatten(values, spec)
+    return len(cache2.key_cache) == len(cache.value_cache)
+
+
 if pv.Version(transformers.__version__) > pv.Version("4.49.99999"):
 
     def make_dynamic_cache(