qnn end to end flow for stories model

Pull Request resolved: #3038

Patch a few changes including:
- support bool tensor type
- support fp16 and fix the 8w8a quantization.
- add two non-supported ops (slice_scatter and index_put) in common_defs.py

stories model working end to end:
AOT:
fp16:
```
python -m examples.models.llama2.export_llama -kv --qnn -c stories110M.pt -p params.json
```
quantize:
```
python -m examples.models.llama2.export_llama -kv --qnn --pt2e_quantize qnn_8a8w -c stories110M.pt -p params.json
```

Runtime:
```
/llama_main --model_path=llama2_fp16_qnn_2.21.pte  --tokenizer_path=tokenizer.bin --prompt="Once"
```
Output:
```
Once upon a time, there was a little girl named Lily. She loved to play outside and explore the world around her. One day, she went on a walk with her mommy and they found a beautiful landscape with lots of trees and flowers.
Lily said, "Mommy, this place is so pretty! Can we take a picture?"
Mommy replied, "Of course, Lily! Let's take a picture to remember the original place we found."
After they took the picture, they continued their walk and saw a bird flying in the sky. Lily said, "MomPyTorchObserver {"prompt_tokens":2,"generated_tokens":125,"model_load_start_ms":1713226585936,"model_load_end_ms":1713226586909,"inference_start_ms":1713226586909,"inference_end_ms":1713226590363,"prompt_eval_end_ms":1713226586966,"first_token_ms":1713226586994,"aggregate_sampling_time_ms":23,"SCALING_FACTOR_UNITS_PER_SECOND":1000}
I 00:00:04.436699 executorch:runner.cpp:414] 	Prompt Tokens: 2    Generated Tokens: 125
I 00:00:04.436703 executorch:runner.cpp:420] 	Model Load Time:		0.973000 (seconds)
I 00:00:04.436732 executorch:runner.cpp:430] 	Total inference time:		3.454000 (seconds)		 Rate: 	36.189925 (tokens/second)
I 00:00:04.436735 executorch:runner.cpp:438] 		Prompt evaluation:	0.057000 (seconds)		 Rate: 	35.087719 (tokens/second)
I 00:00:04.436739 executorch:runner.cpp:449] 		Generated 125 tokens:	3.397000 (seconds)		 Rate: 	36.797174 (tokens/second)
I 00:00:04.436742 executorch:runner.cpp:457] 	Time to first generated token:	0.085000 (seconds)
I 00:00:04.436744 executorch:runner.cpp:464] 	Sampling time over 127 tokens:	0.023000 (seconds)
[INFO] [Qnn ExecuTorch]: Destroy Qnn backend parameters
[INFO] [Qnn ExecuTorch]: Destroy Qnn context
```

Stories model is too small and sensitive to qunatization.
ghstack-source-id: 223091081
@exported-using-ghexport

Differential Revision: [D56119738](https://our.internmc.facebook.com/intern/diff/D56119738/)

Author: cccclai

backends/qualcomm/builders/node_visitor.py

@@ -29,6 +29,7 @@
     QNN_uint16: PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_UFIXED_POINT_16,
 }
 QNN_TENSOR_TYPE_MAP = {
+    torch.bool: PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_BOOL_8,
     torch.float32: PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_FLOAT_32,
     torch.int8: PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_INT_8,
     torch.int16: PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_INT_16,

backends/qualcomm/partition/common_defs.py

@@ -13,6 +13,8 @@
     exir_ops.edge.aten.clone.default,
     exir_ops.edge.aten.index.Tensor,
     exir_ops.edge.aten.full.default,
+    exir_ops.edge.aten.slice_scatter.default,
+    exir_ops.edge.aten.index_put.default,
 ]
 
 allow_list_operator = [

examples/models/llama2/export_llama_lib.py

@@ -19,6 +19,7 @@
 
 import pkg_resources
 import torch
+import torch.nn.functional as F
 from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
 from executorch.backends.xnnpack.partition.xnnpack_partitioner import (
     XnnpackDynamicallyQuantizedPartitioner,
@@ -354,6 +355,13 @@ def build_args_parser() -> argparse.ArgumentParser:
     parser.add_argument(
         "--pt2e_quantize",
         default=None,
+        choices=[
+            "xnnpack_dynamic",
+            "xnnpack_dynamic_qc4",
+            "qnn_8a8w",
+            "qnn_16a16w",
+            "qnn_16a4w",
+        ],
         help="Use PT2E quantization. Comma separated options. e.g. xnnpack_dynamic (for per channel 8 bit weight), xnnpack_dynamic_qc4 (for per channel 4 bit weight), embedding.",
     )
     parser.add_argument(
@@ -624,6 +632,9 @@ def _prepare_for_llama_export(modelname: str, args) -> LlamaEdgeManager:
     if args.use_sdpa_with_kv_cache:
         transforms.append(replace_sdpa_with_custom_op)
 
+    if args.qnn and args.use_kv_cache:
+        transforms.append(replace_sdpa_with_simple_sdpa)
+        transforms.append(replace_causal_mask)
     return (
         load_llama_model(
             checkpoint=checkpoint_path,
@@ -646,13 +657,16 @@ def _export_llama(modelname, args) -> str:  # noqa: C901
     # export_to_edge
     pt2e_quant_params = _get_pt2e_quantization_params(args)
     quantizers = get_pt2e_quantizers(pt2e_quant_params, args)
-    if args.qnn:
-        assert (
-            args.quantization_mode is None
-        ), "Currently qnn backend only supports QnnQuantizer via pt2e flow"
+    quant_dtype = None
+    if args.qnn and args.pt2e_quantize:
         try:
             # pyre-ignore: Undefined import [21]: Could not find a module corresponding to import `executorch.backends.qualcomm.quantizer.quantizer`
-            from executorch.backends.qualcomm.quantizer.quantizer import QnnQuantizer
+            from executorch.backends.qualcomm.quantizer.quantizer import (
+                get_16a4w_qnn_ptq_config,
+                get_default_16bit_qnn_ptq_config,
+                QnnQuantizer,
+                QuantDtype,
+            )
 
             # reset quantizers and pt2e_quant_params from xnnpack backend
             pt2e_quant_params = None
@@ -662,10 +676,36 @@ def _export_llama(modelname, args) -> str:  # noqa: C901
                 "Please install the Qualcomm backend follwing https://pytorch.org/executorch/main/build-run-qualcomm.html"
             )
 
+        backend, quant_config = args.pt2e_quantize.split("_")
+        assert (
+            backend == "qnn"
+        ), f"The quantization config is for backend {backend} instead of qnn."
         # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `qualcomm`.
         qnn_quantizer = QnnQuantizer()
         # more custom quantization are supported including 16a4w etc. default to 8bit quantized
         custom_annotations = ()
+        if quant_config == "8a8w":
+            quant_dtype = QuantDtype.use_8a8w
+            pass
+        elif quant_config == "16a16w":
+            quant_dtype = QuantDtype.use_16a16w
+            qnn_quantizer.add_16bit_quant_ops(qnn_quantizer.SUPPORTED_OPS)
+            qnn_quantizer.set_bit16_op_quant_config(get_default_16bit_qnn_ptq_config())
+        elif quant_config == "16a4w":
+            quant_dtype = QuantDtype.use_16a4w
+            qnn_quantizer.add_16bit_quant_ops(qnn_quantizer.SUPPORTED_OPS)
+            qnn_quantizer.set_bit16_op_quant_config(get_16a4w_qnn_ptq_config())
+            qnn_quantizer.set_per_channel_weight_dtype(
+                weight_dtype_for_16bit_act="int4"
+            )
+        else:
+            raise AssertionError(
+                f"No support for quant type {quant_config}. Support 8a8w, 16a16w and 16a4w."
+            )
+
+        assert (
+            args.quantization_mode is None
+        ), "Currently qnn backend only supports QnnQuantizer via pt2e flow"
         qnn_quantizer.add_custom_quant_annotations(custom_annotations)
         quantizers.append(qnn_quantizer)
 
@@ -781,24 +821,34 @@ def _export_llama(modelname, args) -> str:  # noqa: C901
             )
 
         # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `qualcomm`
-        backend_options = generate_htp_compiler_spec(use_fp16=False)
+        use_fp16 = True
+        skip_node_op_set = {}
+        if args.pt2e_quantize:
+            use_fp16 = False
+            # TODO: fix the lowering error without skipping nodes
+            if quant_dtype == QuantDtype.use_8a8w:
+                raise NotImplementedError("8a8w for llama is still under development")
+            elif quant_dtype == QuantDtype.use_16a16w:
+                raise NotImplementedError("16a16w for llama is still under development")
+            elif quant_dtype == QuantDtype.use_16a4w:
+                raise NotImplementedError("16a4w for llama is still under development")
         partitioners.append(
             # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `qualcomm`
             QnnPartitioner(
                 # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `qualcomm`
                 generate_qnn_executorch_compiler_spec(
                     # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `qualcomm`.
                     soc_model=QcomChipset.SM8650,  # default to SM8650
-                    backend_options=backend_options,
+                    backend_options=generate_htp_compiler_spec(use_fp16=use_fp16),
                     debug=False,
                     saver=False,
                 ),
                 skip_node_id_set={},
-                skip_node_op_set={},
+                skip_node_op_set=skip_node_op_set,
             )
         )
         # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `qualcomm`
-        _transform(builder_exported_to_edge.export_program())
+        _transform(builder_exported_to_edge.edge_manager.exported_program())
 
     if args.generate_etrecord:
         if not builder_exported_to_edge.edge_manager: