fix(generation): Enforce the auto-compile cache check for encoder-decoder models

src/transformers/generation/utils.py

@@ -2133,8 +2133,11 @@ def _valid_auto_compile_criteria(
             generation_config.compile_config is not None and generation_config.compile_config._compile_all_devices
         )
         # Note: for some models that only use linear attention (e.g. Mamba), even a DynamicCache is compilable since all
+        # Encoder-decoder models hold that cache in a subcache, so we unwrap it to check the cache the decoder actually generates with
+        decoder_cache = cache.self_attention_cache if isinstance(cache, EncoderDecoderCache) else cache
+        # Note: for some models that only use linear attention (e.g. Mamba), even a DynamicCache is compilable since all
         # layers are, but we don't want to ALWAYS compile when calling `generate`, so we check the type
-        using_compilable_cache = cache is not None and cache.is_compileable and type(cache) is not DynamicCache
+        using_compilable_cache = cache is not None and cache.is_compileable and type(decoder_cache) is not DynamicCache
         can_compile = valid_hardware and using_compilable_cache
 
         # Exception 1: Some quantization methods do not support compilation

tests/generation/test_utils.py

@@ -5114,6 +5114,23 @@ def test_cpu_offload_doesnt_compile(self):
         _ = model_cpu.generate(input_ids, **generate_kwargs)
         self.assertFalse(hasattr(model_cpu, "_compiled_call"))
 
+    def test_compileable_default_cache_doesnt_compile_encoder_decoder(self):
+        """Test that a compileable default cache doesn't trigger compilation on encoder-decoder models either"""
+        model = AutoModelForSeq2SeqLM.from_pretrained("hf-internal-testing/tiny-random-bart")
+        decoder_config = model.config.get_text_config(decoder=True)
+        # Linear attention layers are statically shaped, so the default `DynamicCache` is compileable (e.g. Mamba)
+        decoder_config.layer_types = ["linear_attention"] * decoder_config.num_hidden_layers
+        self_attention_cache = DynamicCache(config=decoder_config)
+        self.assertTrue(self_attention_cache.is_compileable)
+
+        cache = EncoderDecoderCache(
+            self_attention_cache, DynamicCache(config=model.config.get_text_config(decoder=True))
+        )
+        generation_config = GenerationConfig()
+        generation_config.compile_config = CompileConfig()
+        generation_config.compile_config._compile_all_devices = True  # force compilation (e.g. fast CI, CPU)
+        self.assertFalse(model._valid_auto_compile_criteria({"past_key_values": cache}, generation_config))
+
     def test_custom_generate_from_argument_in_generate(self):
         """Tests that the `custom_generate` argument is used when passed to `generate`"""
         model = AutoModelForCausalLM.from_pretrained(