Cleaner Cache `dtype` and `device` extraction for CUDA graph generation for...

Cleaner Cache `dtype` and `device` extraction for CUDA graph generation for quantizers compatibility (#29079)

* input_layernorm as the beacon of hope

* cleaner dtype extraction

* AQLM + CUDA graph test

* is available check

* shorter text test

src/transformers/models/llama/modeling_llama.py

@@ -817,9 +817,13 @@ class LlamaPreTrainedModel(PreTrainedModel):
             self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
 
         for layer in self.model.layers:
-            weights = layer.self_attn.o_proj.weight
+            device = layer.input_layernorm.weight.device
+            if hasattr(self.config, "_pre_quantization_dtype"):
+                dtype = self.config._pre_quantization_dtype
+            else:
+                dtype = layer.self_attn.o_proj.weight.dtype
             layer.self_attn.past_key_value = cache_cls(
-                self.config, max_batch_size, max_cache_len, device=weights.device, dtype=weights.dtype
+                self.config, max_batch_size, max_cache_len, device=device, dtype=dtype
             )
 
     def _reset_cache(self):

tests/quantization/aqlm_integration/test_aqlm.py

@@ -14,10 +14,13 @@
 # limitations under the License.
 
 import gc
+import importlib
 import tempfile
 import unittest
 
-from transformers import AqlmConfig, AutoConfig, AutoModelForCausalLM, AutoTokenizer, OPTForCausalLM
+from packaging import version
+
+from transformers import AqlmConfig, AutoConfig, AutoModelForCausalLM, AutoTokenizer, OPTForCausalLM, StaticCache
 from transformers.testing_utils import (
     require_accelerate,
     require_aqlm,
@@ -26,7 +29,7 @@ from transformers.testing_utils import (
     slow,
     torch_device,
 )
-from transformers.utils import is_accelerate_available, is_torch_available
+from transformers.utils import is_accelerate_available, is_aqlm_available, is_torch_available
 
 
 if is_torch_available():
@@ -71,11 +74,12 @@ class AqlmConfigTest(unittest.TestCase):
 @require_aqlm
 @require_accelerate
 class AqlmTest(unittest.TestCase):
-    model_name = "BlackSamorez/Mixtral-8x7b-AQLM-2Bit-1x16-hf-test-dispatch"
+    model_name = "BlackSamorez/Llama-2-7b-AQLM-2Bit-1x16-hf"
 
     input_text = "Hello my name is"
+    max_new_tokens = 32
 
-    EXPECTED_OUTPUT = "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am currently a sophomore and am majoring in Psychology. I am"
+    EXPECTED_OUTPUT = "Hello my name is Katie. I am a 20 year old college student. I am a very outgoing person. I love to have fun and be active. I"
 
     device_map = "cuda"
 
@@ -144,7 +148,7 @@ class AqlmTest(unittest.TestCase):
         """
         input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)
 
-        output = self.quantized_model.generate(**input_ids, max_new_tokens=40)
+        output = self.quantized_model.generate(**input_ids, max_new_tokens=self.max_new_tokens)
         self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
 
     def test_raise_if_non_quantized(self):
@@ -164,7 +168,7 @@ class AqlmTest(unittest.TestCase):
 
             input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)
 
-            output = model.generate(**input_ids, max_new_tokens=40)
+            output = model.generate(**input_ids, max_new_tokens=self.max_new_tokens)
             self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
 
     @require_torch_multi_gpu
@@ -178,6 +182,56 @@ class AqlmTest(unittest.TestCase):
 
         self.assertTrue(set(quantized_model.hf_device_map.values()) == {0, 1})
 
-        output = quantized_model.generate(**input_ids, max_new_tokens=40)
+        output = quantized_model.generate(**input_ids, max_new_tokens=self.max_new_tokens)
 
         self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
+
+    @unittest.skipUnless(
+        is_aqlm_available() and version.parse(importlib.metadata.version("aqlm")) >= version.parse("1.0.3"),
+        "test requires `aqlm>=1.0.3`",
+    )
+    def test_quantized_model_compile(self):
+        """
+        Simple test that checks if the quantized model is working properly
+        """
+
+        # Sample tokens greedily
+        def decode_one_tokens(model, cur_token, input_pos, cache_position):
+            logits = model(
+                cur_token, position_ids=input_pos, cache_position=cache_position, return_dict=False, use_cache=True
+            )[0]
+            new_token = torch.argmax(logits[:, [-1]], dim=-1).to(torch.int)
+
+            return new_token
+
+        # Tokenize the test input
+        input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)["input_ids"]
+        seq_length = input_ids.shape[1]
+
+        # Setup static KV cache for generation
+        self.quantized_model._setup_cache(StaticCache, 1, max_cache_len=seq_length + self.max_new_tokens + 1)
+
+        # Allocate token ids to be generated and copy prefix ids
+        cache_position = torch.arange(seq_length, device=torch_device)
+        generated_ids = torch.zeros(1, seq_length + self.max_new_tokens, dtype=torch.int, device=torch_device)
+        generated_ids[:, cache_position] = input_ids.to(torch_device).to(torch.int)
+
+        # Do a forward pass to fill the prefix cache and compile the kernels if necessary
+        logits = self.quantized_model(input_ids, cache_position=cache_position, return_dict=False, use_cache=True)[0]
+        next_token = torch.argmax(logits[:, [-1]], dim=-1).to(torch.int)
+        generated_ids[:, [seq_length]] = next_token
+
+        with torch.no_grad():
+            # Compile the CUDA graph
+            decode_one_tokens = torch.compile(decode_one_tokens, mode="reduce-overhead", fullgraph=True)
+
+            # Generate tokens one by one
+            cache_position = torch.tensor([seq_length + 1], device=torch_device)
+            for _ in range(1, self.max_new_tokens):
+                with torch.backends.cuda.sdp_kernel(enable_flash=False, enable_mem_efficient=False, enable_math=True):
+                    next_token = decode_one_tokens(self.quantized_model, next_token.clone(), None, cache_position)
+                    generated_ids.index_copy_(1, cache_position, next_token)
+                cache_position += 1
+
+        # Check generated text
+        self.assertEqual(self.tokenizer.decode(generated_ids[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)