Support Mistral

MODELS.md

@@ -121,6 +121,10 @@ Does not work with flexible sequence length and therefore does not support `use_
 - ✅ GPT2ForTokenClassification
 - ⚠️ GPT2LMHeadModel (no `use_past`)
 
+**Llama**
+
+- ✅ LlamaForCausalLM
+
 **M2M100**
 
 - ⚠️ M2M100Model (currently supports only `use_past=False`)
@@ -132,6 +136,10 @@ Does not work with flexible sequence length and therefore does not support `use_
 - ? MarianForCausalLM
 - ⚠️ MarianMTModel (currently supports only `use_past=False`)
 
+**Mistral**
+
+- ✅ MistralForCausalLM
+
 **MobileBERT**
 
 - ✅ MobileBertModel

src/exporters/coreml/features.py

@@ -291,6 +291,12 @@ class FeaturesManager:
             "text2text-generation",
             coreml_config_cls="models.marian.MarianMTCoreMLConfig",
         ),
+        "mistral": supported_features_mapping(
+            "feature-extraction",
+            "text-generation",
+            "text-classification",
+            coreml_config_cls="models.mistral.MistralCoreMLConfig",
+        ),
         "mobilebert": supported_features_mapping(
             "feature-extraction",
             "fill-mask",

src/exporters/coreml/models.py

@@ -282,6 +282,24 @@ class MarianMTCoreMLConfig(CoreMLConfig):
     modality = "text"
 
 
+class MistralCoreMLConfig(CoreMLConfig):
+    modality = "text"
+
+    def patch_pytorch_ops(self):
+        # Workaround for https://github.com/apple/coremltools/pull/2017
+        def log(context, node):
+            from coremltools.converters.mil import Builder as mb
+            from coremltools.converters.mil.mil import types
+
+            a = context[node.inputs[0]]
+            if types.is_int(a.dtype):
+                a = mb.cast(x=a, dtype="fp32")
+            x = mb.log(x=a, name=node.name)
+            context.add(x)
+
+        return {"log": log}
+
+
 class MobileBertCoreMLConfig(CoreMLConfig):
     modality = "text"