[NPU] Add Optimized Support for Llama3.2-1B/3B on NPU (#12339)

* Add initial support for llama3.2-1b/3b

* move llama3.2 support into current llama_mp impl

python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md

@@ -7,6 +7,8 @@ In this directory, you will find examples on how to directly run HuggingFace `tr
 |------------|----------------------------------------------------------------|
 | Llama2 | [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) |
 | Llama3 | [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) |
+| Llama3.2-1B | [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) |
+| Llama3.2-3B | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
 | Chatglm3 | [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) |
 | Chatglm2 | [THUDM/chatglm2-6b](https://huggingface.co/THUDM/chatglm2-6b) |
 | Qwen2 | [Qwen/Qwen2-7B-Instruct](https://huggingface.co/Qwen/Qwen2-7B-Instruct), [Qwen/Qwen2-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2-1.5B-Instruct) |
@@ -33,6 +35,9 @@ conda activate llm
 
 :: install ipex-llm with 'npu' option
 pip install --pre --upgrade ipex-llm[npu]
+
+:: [optional] for Llama-3.2-1B-Instruct & Llama-3.2-3B-Instruct
+pip install transformers==4.45.0 accelerate==0.33.0
 ```
 
 ## 2. Runtime Configurations
@@ -82,6 +87,8 @@ done
 The examples below show how to run the **_optimized HuggingFace model implementations_** on Intel NPU, including
 - [Llama2-7B](./llama.py)
 - [Llama3-8B](./llama.py)
+- [Llama3.2-1B](./llama.py)
+- [Llama3.2-3B](./llama.py)
 - [Qwen2-1.5B](./qwen.py)
 - [Qwen2.5-7B](./qwen.py)
 - [MiniCPM-1B](./minicpm.py)
@@ -106,6 +113,12 @@ python llama.py
 :: to run Meta-Llama-3-8B-Instruct (LNL driver version: 32.0.101.2715)
 python llama.py --repo-id-or-model-path meta-llama/Meta-Llama-3-8B-Instruct
 
+:: to run Llama-3.2-1B-Instruct
+python llama.py --repo-id-or-model-path meta-llama/Llama-3.2-1B-Instruct
+
+:: to run Llama-3.2-3B-Instruct
+python llama.py --repo-id-or-model-path meta-llama/Llama-3.2-3B-Instruct
+
 :: to run Qwen2-1.5B-Instruct (LNL driver version: 32.0.101.2715)
 python qwen.py
 
@@ -145,6 +158,12 @@ python llama.py --disable-transpose-value-cache
 :: to run Meta-Llama-3-8B-Instruct (LNL driver version: 32.0.101.2715)
 python llama.py --repo-id-or-model-path meta-llama/Meta-Llama-3-8B-Instruct --disable-transpose-value-cache
 
+:: to run Llama-3.2-1B-Instruct
+python llama.py --repo-id-or-model-path meta-llama/Llama-3.2-1B-Instruct --disable-transpose-value-cache
+
+:: to run Llama-3.2-3B-Instruct
+python llama.py --repo-id-or-model-path meta-llama/Llama-3.2-3B-Instruct --disable-transpose-value-cache
+
 :: to run Qwen2-1.5B-Instruct (LNL driver version: 32.0.101.2715)
 python qwen.py --disable-transpose-value-cache
 

python/llm/example/NPU/HF-Transformers-AutoModels/README.md

@@ -10,6 +10,8 @@ This folder contains examples of running IPEX-LLM on Intel NPU:
 |------------|----------------------------------------------------------------|
 | Llama2 | [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) |
 | Llama3 | [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) |
+| Llama3.2-1B | [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) |
+| Llama3.2-3B | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
 | Chatglm3 | [THUDM/chatglm3-6b](https://huggingface.co/THUDM/chatglm3-6b) |
 | Chatglm2 | [THUDM/chatglm2-6b](https://huggingface.co/THUDM/chatglm2-6b) |
 | Qwen2 | [Qwen/Qwen2-7B-Instruct](https://huggingface.co/Qwen/Qwen2-7B-Instruct), [Qwen/Qwen2-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2-1.5B-Instruct) |

python/llm/src/ipex_llm/transformers/npu_models/convert_mp.py

@@ -173,7 +173,8 @@ def convert_llama(
         intra_pp=None,
         transpose_value_cache=True,
 ):
-    from ipex_llm.transformers.npu_models.llama_mp import gen_llama_fused_model_forward
+    from ipex_llm.transformers.npu_models.llama_mp import gen_llama_fused_model_forward,\
+        gen_llama_32_fused_model_forward
     from ipex_llm.transformers.npu_models.llama_mp import DecodeRunner, PrefillRunner
     from transformers.models.llama.modeling_llama import LlamaModel
 
@@ -193,9 +194,18 @@ def convert_llama(
         max_prompt_len=max_prompt_len,
         transpose_value_cache=transpose_value_cache,
     )
-    llama_model_forward = gen_llama_fused_model_forward(
-        prefill_runner=prefill_runner, decode_runner=decode_runner
-    )
+    from packaging import version
+    import transformers
+    trans_version = transformers.__version__
+    if version.parse(trans_version) == version.parse("4.45.0"):
+        # llama-3.2-3B & llama-3.2-1B
+        llama_model_forward = gen_llama_32_fused_model_forward(
+            prefill_runner=prefill_runner, decode_runner=decode_runner
+        )
+    else:
+        llama_model_forward = gen_llama_fused_model_forward(
+            prefill_runner=prefill_runner, decode_runner=decode_runner
+        )
     convert_forward(model, LlamaModel, llama_model_forward)
     from transformers.models.llama.modeling_llama import LlamaForCausalLM
     from ipex_llm.transformers.npu_models.llama_mp import llama2_casullm_forward

python/llm/src/ipex_llm/transformers/npu_models/kv.py

@@ -145,7 +145,7 @@ class DynamicFusedNormalCache(DynamicCache):
     # Experimental support for fused decoderlayer implementation on NPU
     # Currently only for llama2
 
-    def __init__(self) -> None:
+    def __init__(self, num_hidden_layers: Optional[int] = None) -> None:
         self.key_cache: Dict[int, torch.Tensor] = {}
         self.value_cache: Dict[int, torch.Tensor] = {}
         self.min_layer_idx = sys.maxsize
@@ -158,6 +158,9 @@ def update(
         cache_kwargs: Optional[Dict[str, Any]]=None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
 
+        if key_states == []:
+            return key_states, value_states
+
         batch_size, num_heads, seq_len, head_dim = key_states.shape
 
         max_seq_length = cache_kwargs["max_seq_len"] if "max_seq_len" in cache_kwargs else None