Qualcomm AI Engine Direct - Suport batch prefill mode for llama3.2 (#6983)

* Qualcomm AI Engine Direct - Suport bert mode for llama3.2

- Enable bert mode
- Change input sequence of static_llama
- Tag bert output as uint8
- Unify both 1b and 3b in 1 runner
- Add hybrid IO memory for llama3_2 runner
- Align timer with llama

* Rebase and minor fix

- Fix rebase conflict
- Change input dtype of calibration function

* Change bert to batch prefill

* Fix compile error

* Fix lint

- Fix transformers version
- Refine pass quantization tagging function
- Rebase

* Add one line in the end of CmakeList

* Remove trailing line of CmakeList

* Move noqa to correct line number

---------

Co-authored-by: Joey Tsai <[email protected]>

Author: chunit-quic

backends/qualcomm/quantizer/custom_annotation.py

@@ -22,7 +22,9 @@
 from torch.fx import Node
 
 
-def annotate_matmul_16a8w(gm: torch.fx.GraphModule) -> None:  # noqa: C901
+def annotate_matmul_16a8w(  # noqa: C901
+    gm: torch.fx.GraphModule, traverse_input1=True
+) -> None:
     """
     This function is specific for matmul op 16a8w.
     """
@@ -99,7 +101,8 @@ def annotate_matmul_input1(node: Node):
     for node in gm.graph.nodes:
         if node.op == "call_function" and node.target == torch.ops.aten.matmul.default:
             annotate_matmul(node, quantization_config_16a8w)
-            annotate_matmul_input1(node.args[1])
+            if traverse_input1:
+                annotate_matmul_input1(node.args[1])
 
 
 def custom_annotate_llama_matmul_16a8w(gm: torch.fx.GraphModule) -> None:  # noqa: C901

examples/qualcomm/oss_scripts/llama2/llama.py

@@ -177,14 +177,15 @@ def annotate_conv2d(node: Node, quantization_config: QuantizationConfig) -> None
                     )
 
 
-def calibrate(
+def _kv_calibrate(
     example_inputs,
     user_prompts,
     module: torch.fx.GraphModule,
     tokenizer_model_path="tokenizer.model",
+    max_seq_len=512,
 ):
     sp_model = SentencePieceProcessor(model_file=tokenizer_model_path)
-    _, _, atten_mask, k_caches, v_caches = example_inputs
+    _, atten_mask, _, k_caches, v_caches = example_inputs
 
     # TODO: change criteria & support batch inputs if necessary
     pos = torch.tensor(0, dtype=torch.int32)
@@ -202,11 +203,11 @@ def sample_top_p(probs: torch.Tensor, top_p: float) -> torch.Tensor:
         return probs_indices.gather(dim=-1, index=next_token)
 
     with torch.no_grad():
-        while token_list[-1] != sp_model.eos_id() and pos < 128:
+        while token_list[-1] != sp_model.eos_id() and pos < max_seq_len - 1:
             logits, new_k_caches, new_v_caches = module(
                 torch.full((1, 1), token_list[pos]),
-                torch.full((1, 1), pos),
                 atten_mask,
+                torch.full((1, 1), pos),
                 *k_caches,
                 *v_caches,
             )
@@ -228,15 +229,84 @@ def sample_top_p(probs: torch.Tensor, top_p: float) -> torch.Tensor:
     print(f"calibration data:\n{sp_model.decode(token_list)}")
 
 
+def _batch_prefill_calibrate(
+    example_inputs,
+    user_prompts,
+    module: torch.fx.GraphModule,
+    tokenizer_model_path="tokenizer.model",
+    max_seq_len=512,
+):
+    sp_model = SentencePieceProcessor(model_file=tokenizer_model_path)
+    _, atten_mask = example_inputs
+    max_cache_len = max_seq_len - 1
+
+    # TODO: change criteria & support batch inputs if necessary
+    token_list = sp_model.encode(user_prompts, bos=True, eos=False)
+    token_list = torch.tensor(token_list)[:max_cache_len].reshape(1, -1)
+    last_prompt_pos = token_list.numel()
+    if last_prompt_pos < max_cache_len:
+        token_list = torch.cat(
+            [
+                token_list,
+                torch.zeros((1, max_cache_len - last_prompt_pos), dtype=torch.int32),
+            ],
+            dim=1,
+        )
+    else:
+        token_list = token_list[:, :max_cache_len]
+
+    with torch.no_grad():
+        logits, new_k_caches, new_v_caches = module(
+            token_list,
+            atten_mask,
+        )
+        predict = [torch.argmax(logits[:, last_prompt_pos - 1], dim=-1).item()]
+
+    print(f"calibration data:\n{sp_model.decode(predict)}")
+
+
+def calibrate(
+    example_inputs,
+    user_prompts,
+    module: torch.fx.GraphModule,
+    tokenizer_model_path="tokenizer.model",
+    max_seq_len=512,
+):
+    if len(example_inputs) == 2:
+        _batch_prefill_calibrate(
+            example_inputs,
+            user_prompts,
+            module,
+            tokenizer_model_path,
+            max_seq_len,
+        )
+    elif len(example_inputs) == 5:
+        _kv_calibrate(
+            example_inputs,
+            user_prompts,
+            module,
+            tokenizer_model_path,
+            max_seq_len,
+        )
+    else:
+        raise RuntimeError("Get wrong inputs")
+
+
 class SingleLlama:
     def __init__(self, llama_model) -> None:
         super().__init__()
         self.llama_model = llama_model
         self.quant_dtype = None
         self.llama_meta = self.llama_model.get_metadata()
         self.has_quant_io = False
-        tokens, pos_ids, atten_mask, k_caches, v_caches = self.get_example_inputs()
-        self.inputs = (tokens, pos_ids, atten_mask, *k_caches, *v_caches)
+        if self.llama_meta["get_use_kv_cache"]:
+            tokens, atten_mask, pos_ids, k_caches, v_caches = self.get_example_inputs(
+                use_kv_cache=True
+            )
+            self.inputs = (tokens, atten_mask, pos_ids, *k_caches, *v_caches)
+        else:
+            tokens, atten_mask = self.get_example_inputs(use_kv_cache=False)
+            self.inputs = (tokens, atten_mask)
 
     def _tag_kv_ios(self, gm: torch.fx.GraphModule, kv_type):
         if not self.has_quant_io:
@@ -256,11 +326,17 @@ def _tag_kv_ios(self, gm: torch.fx.GraphModule, kv_type):
                 n.meta[QCOM_QUANTIZED_IO] = kv_type
             elif n.op == "output":
                 for a in n.args[0]:
+                    # single head, kv mode
                     if (
                         a.meta["val"].flatten().size()[0]
                         == self.llama_meta["get_head_dim"]
                     ):
                         a.meta[QCOM_QUANTIZED_IO] = kv_type
+                    # single head, batch_prefill mode
+                    elif a.meta["val"].flatten().size()[0] == self.llama_meta[
+                        "get_head_dim"
+                    ] * (self.llama_meta["get_max_seq_len"] - 1):
+                        a.meta[QCOM_QUANTIZED_IO] = kv_type
 
     def quantize(self, quant_dtype, custom_annotations=()):
         self.quant_dtype = quant_dtype
@@ -281,11 +357,13 @@ def quantize(self, quant_dtype, custom_annotations=()):
             ).module()
             fx_graph_module = prepare_pt2e(fx_graph_module, quantizer)
         print("Quantizing the model...")
+
         calibrate(
-            self.get_example_inputs(),
+            self.get_example_inputs(self.llama_meta["get_use_kv_cache"]),
             args.prompt,
             fx_graph_module,
             tokenizer_model_path=args.tokenizer_model,
+            max_seq_len=args.seq_len,
         )
 
         self.llama_model = convert_pt2e(fx_graph_module)
@@ -328,18 +406,29 @@ def lowering_modules(
             with open(f"{work_space}/{pte_filename}.pte", "wb") as file:
                 exec_prog_mgr.write_to_file(file)
 
-    def get_example_inputs(self):
-        return self.llama_model.get_example_inputs()
+    def get_example_inputs(self, use_kv_cache=True):
+        return self.llama_model.get_example_inputs(use_kv_cache)
 
 
 def compile(args):
     os.makedirs(args.artifact, exist_ok=True)
     start_ts = time.time()
+
+    if args.model_mode == "kv":
+        use_kv_cache = output_new_cache_only = True
+    elif args.model_mode == "batch_prefill" or args.model_mode == "hybrid":
+        raise NotImplementedError(
+            f"model_mode {args.model_mode} is not implemented yet."
+        )
+    else:
+        raise RuntimeError(f"No such model_mode {args.model_mode}.")
+
     with open(args.params) as f:
         config = ModelArgs(**json.load(f))
         # TODO: support batch inputs if necessary
         config.max_batch_size = 1
-        config.max_seq_len = 1024
+        config.max_seq_len = args.seq_len
+        config.use_kv_cache = use_kv_cache
     state_dict = torch.load(
         args.checkpoint, weights_only=True, map_location="cpu", mmap=True
     )
@@ -348,7 +437,7 @@ def compile(args):
 
     llama_instance = None
     with torch.device("meta"):
-        llama_instance = LlamaModel(config, output_new_cache_only=True)
+        llama_instance = LlamaModel(config, output_new_cache_only=output_new_cache_only)
     if "model" in state_dict:
         state_dict = state_dict["model"]
     llama_instance.load_state_dict(
@@ -398,6 +487,11 @@ def compile(args):
 def inference(args, pre_gen_pte=""):
     workspace = f"/data/local/tmp/{getpass.getuser()}/executorch/single_llama"
 
+    if args.model_mode != "kv":
+        raise NotImplementedError(
+            f"model_mode {args.model_mode} is not implemented yet."
+        )
+
     runner_args = " ".join(
         [
             f"--model_path {pte_filename}.pte",
@@ -550,6 +644,21 @@ def post_process():
         type=str,
     )
 
+    parser.add_argument(
+        "--num_sharding",
+        type=int,
+        default=0,
+        help="Specify the number of splits by inserting the fallback custom op. The graph will be split evenly by layers.",
+    )
+
+    parser.add_argument(
+        "--model_mode",
+        help="Export and inference batch_prefill mode, kv mode or hybrid(TBD) mode",
+        default="kv",
+        choices=["batch_prefill", "kv", "hybrid"],
+        type=str,
+    )
+
     args = parser.parse_args()
     if args.compile_only and args.pre_gen_pte:
         exit("Cannot set both compile_only and pre_gen_pte as true")