Qualcomm AI Engine Direct - Improve GA Static Phi-4-mini accuracy (#13573)

Summary:

- Refactor custom annotation for R3
- Fix warning message in quantization
- Add phi-4-mini setting into README
- Fixed segmemtation fault when run the model with sharding
- Add a test case for phi-4 in test_qnn_delegate.py
- Add new parameter "group_size" in llama.py to set block size in block
quantization

## Sample Script
```
python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -s ${SERIAL_NUM} -m ${SOC_MODEL} \ 
--ptq 16a4w_block --group_size 16 --checkpoint consolidated.00.pth --params params.json --num_sharding 4 \
--tokenizer_model tokenizer.model --decoder_model phi_4_mini --model_mode hybrid --prefill_ar_len 128 \ 
--max_seq_len 1024 --prompt "I would like to learn python, could you teach me with a simple example?"
```

## Result
Stats with QNN2.37.0 on SM8750
Accuracy: 10.82
Token Rate: 22.727273
Results:
--prompt "I would like to learn python, could you teach me with a simple
example?"
```
<|user|>I would like to learn python, could you teach me with one simple program?<|end|><|assistant|>Of course! Let's get started with a simple Python program. We'll create a simple program that asks for your name and then greets you.

```python
# Ask for the user's name
name = input("Please enter your name: ")

# Greet the user
print(f"Hello, {name}! Welcome to the world of Python.")
```

To run this program, you would need to copy the code into a Python environment (like an IDE or a Python interpreter). When you run the program, it will prompt you to enter your name, and then it will greet you by name. Enjoy learning Python!<|end|>
```

## Test plan
Added E2E test to test_qnn_delegate.py

cc: @haowhsu-quic

Author: shewu-quic

backends/qualcomm/builders/node_visitor.py

@@ -162,7 +162,7 @@ def make_qnn_per_block_config(self, node: torch.fx.Node, quant_attrs: Dict):
         for ch in range(num_channels):
             max_scale = scales[ch].reshape(1, -1).amax(dim=-1) / num_steps
             q_scales = torch.clamp(
-                input=scales[ch] / max_scale,
+                input=torch.round(input=scales[ch] / max_scale),
                 min=1,
                 max=2**bitwidth_of_scale,
             ).to(quant_scales_dtype)

backends/qualcomm/quantizer/custom_annotation.py

@@ -317,6 +317,7 @@ def annotate_matmul_input1(node: Node, is_qat: str):
                 torch.ops.aten.transpose.int,
                 torch.ops.aten.view.default,
                 torch.ops.aten.reshape.default,
+                torch.ops.aten.slice.Tensor,
             ]:
                 annotate_single_in_single_out(node, quantization_config_8a8w)
                 node = node.args[0]
@@ -340,7 +341,11 @@ def annotate_matmul_input1(node: Node, is_qat: str):
                     node, quantization_config=quantization_config_8a4w_per_channel
                 )
                 break
-            elif node.target in [torch.ops.aten.add.Tensor, torch.ops.aten.sub.Tensor]:
+            elif node.target in [
+                torch.ops.aten.add.Tensor,
+                torch.ops.aten.sub.Tensor,
+                torch.ops.aten.matmul.default,
+            ]:
                 break
             else:
                 print(f"The node ({node}) is not expected in the input1 of the matmul")
@@ -356,7 +361,12 @@ def annotate_matmul_input1(node: Node, is_qat: str):
         )
 
     for node in gm.graph.nodes:
-        if node.op == "call_function" and node.target == torch.ops.aten.matmul.default:
+        if (
+            node.op == "call_function"
+            and node.target == torch.ops.aten.matmul.default
+            and all(arg.op == "call_function" for arg in node.args)
+        ):
+            # Only apply custom annotation on Q @ K^T @ V
             annotate_matmul(node, quantization_config_16a8w)
             annotate_matmul_input1(node.args[1], is_qat=is_qat)
 

backends/qualcomm/runtime/backends/QnnImplementation.cpp

@@ -51,16 +51,8 @@ Error QnnImplementation::StartBackend(
     const std::string& lib_path,
     const QnnSaver_Config_t** saver_config) {
   Qnn_ErrorHandle_t error = QNN_SUCCESS;
-  // RTLD_GLOBAL is needed on x86 as HTP op package has a requirement for the
-  // symbols in backend to be visible. Using RTLD_LOCAL on Android to allow full
-  // unloading of HTP backend shared library on dlclose() as RTLD_GLOBAL isn't
-  // letting it happen.
   void* lib_handle = nullptr;
-#if defined(__ANDROID__)
-  lib_handle = dlopen(lib_path.c_str(), RTLD_NOW | RTLD_LOCAL);
-#else
   lib_handle = dlopen(lib_path.c_str(), RTLD_NOW | RTLD_GLOBAL);
-#endif
   if (lib_handle == nullptr) {
     QNN_EXECUTORCH_LOG_ERROR(
         "Cannot Open QNN library %s, with error: %s",

backends/qualcomm/tests/test_qnn_delegate.py

@@ -1134,7 +1134,8 @@ def test_qnn_backend_where(self):
             (torch.randn(30, 20),),
         ]
         for i, module in enumerate(modules):
-            self.lower_module_and_test_output(module, sample_inputs[i])
+            with self.subTest(i=i):
+                self.lower_module_and_test_output(module, sample_inputs[i])
 
     def test_qnn_backend_masked_fill(self):
         module = MaskedFill()  # noqa: F405
@@ -2571,8 +2572,9 @@ def test_qnn_backend_where(self):
             (torch.randn(30, 20),),
         ]
         for i, module in enumerate(modules):
-            module = self.get_qdq_module(module, sample_inputs[i])
-            self.lower_module_and_test_output(module, sample_inputs[i])
+            with self.subTest(i=i):
+                module = self.get_qdq_module(module, sample_inputs[i])
+                self.lower_module_and_test_output(module, sample_inputs[i])
 
     def test_qnn_backend_masked_fill(self):
         module = MaskedFill()  # noqa: F405
@@ -4541,6 +4543,77 @@ def test_llama_stories_110m(self):
                 if not self.compile_only and not self.enable_x86_64:
                     self.assertGreaterEqual(msg["inference_speed"], 220)  # Lanai
 
+    def test_static_phi4(self):
+        if not self.required_envs():
+            self.skipTest("missing required envs")
+
+        prompt = "My favourite condiment is "
+        cmds = [
+            "python",
+            f"{self.executorch_root}/examples/qualcomm/oss_scripts/llama/llama.py",
+            "--artifact",
+            self.artifact_dir,
+            "--build_folder",
+            self.build_folder,
+            "--model",
+            self.model,
+            "--ip",
+            self.ip,
+            "--port",
+            str(self.port),
+            "--prompt",
+            f"{prompt}",
+            "--ptq",
+            "16a4w_block",
+            "--group_size",
+            "16",
+            "--decoder_model",
+            "phi_4_mini",
+            "--model_mode",
+            "kv",
+            "--max_seq_len",
+            "1024",
+            "--num_sharding",
+            "8",
+            "--eval_perplexity",
+            "--tasks",
+            "wikitext",
+            "--limit",
+            "1",
+        ]
+        if self.compile_only:
+            cmds.extend(["--compile_only"])
+        elif self.device:
+            cmds.extend(["--device", self.device])
+        if self.host:
+            cmds.extend(["--host", self.host])
+        elif self.enable_x86_64:
+            cmds.extend(["--enable_x86_64"])
+        if self.pre_gen_pte:
+            cmds.extend(["--pre_gen_pte", self.pre_gen_pte])
+            cmds.extend(
+                [
+                    "--quant_attrs_path",
+                    f"{self.pre_gen_pte}/kv_llama_qnn_quant_attrs.json",
+                ]
+            )
+
+        p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
+        with Listener((self.ip, self.port)) as listener:
+            conn = listener.accept()
+            p.communicate()
+            msg = json.loads(conn.recv())
+            if "Error" in msg:
+                self.fail(msg["Error"])
+            else:
+                inference_speed_ref = {"SM8650": 14, "SM8750": 19}
+                self.assertLessEqual(msg["wiki_ppl"], 12)
+                self.assertLessEqual(msg["pte_size"], 4000000000)  # 4gb
+                if self.model in inference_speed_ref:
+                    self.assertGreaterEqual(
+                        msg["inference_speed"], inference_speed_ref[self.model]
+                    )
+
     def test_static_qwen2_5(self):
         if not self.required_envs():
             self.skipTest("missing required envs")

examples/qualcomm/oss_scripts/llama/README.md

@@ -69,6 +69,12 @@ Default example using hybrid mode.
 python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -s ${SERIAL_NUM} -m ${SOC_MODEL} --ptq 16a4w --checkpoint consolidated.00.pth --params params.json --tokenizer_model tokenizer.model --decoder_model llama3_2 --model_mode hybrid --prefill_ar_len 32 --max_seq_len 128 --prompt "what is 1+1"
 ```
 
+#### Phi4-mini-instruct
+Default example using hybrid mode.
+```bash
+python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -s ${SERIAL_NUM} -m ${SOC_MODEL} --ptq 16a4w_block --group_size 16 --checkpoint consolidated.00.pth --params params.json --tokenizer_model tokenizer.model --decoder_model phi_4_mini --model_mode hybrid --prefill_ar_len 128 --max_seq_len 1024 --num_sharding 8 --prompt "I would like to learn python, could you teach me with a simple example?"
+```
+
 #### QWEN2.5 0.5B
 Default example using hybrid mode
 ```bash
@@ -99,6 +105,7 @@ Default example using hybrid mode.
 python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -H mlgtw-linux -s ${SERIAL_NUM} -m ${SOC_MODEL} --ptq 16a8w --decoder_model smollm2_135m --model_mode hybrid --prefill_ar_len 128 --max_seq_len 1024 --prompt "I would like to learn python, could you teach me with a simple example?"
 ```
 
+
 ### KV Cache update mechanism
 We have two distinct mechanisms for updating the key-value (KV) cache, which can be selected at runtime. Shift Pointer and Smart Mask.
 

examples/qualcomm/oss_scripts/llama/llama.py

@@ -117,6 +117,8 @@
 FORMAT = "[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s"
 logging.basicConfig(level=logging.INFO, format=FORMAT)
 logging.getLogger().setLevel(logging.INFO)
+# Avoid the error message "Could not initialize NNPACK! Reason: Unsupported hardware."
+torch.backends.nnpack.set_flags(False)
 
 
 def next_power_of_two(n):
@@ -235,10 +237,16 @@ def quantize(
             ).module()
 
             if quant_dtype == QuantDtype.use_16a4w_block:
+                if args.group_size is None:
+                    raise ValueError(
+                        "Group size is required when use quant_dtype 16a4w_block"
+                    )
                 conv_nodes = [
                     n for n in fx_graph_module.graph.nodes if "conv" in n.name
                 ]
-                block_size_map = {n.name: (1, 64, 1, 1) for n in conv_nodes}
+                block_size_map = {
+                    n.name: (1, args.group_size, 1, 1) for n in conv_nodes
+                }
                 quantizer.set_block_size_map(block_size_map)
 
             fx_graph_module = prepare_pt2e(fx_graph_module, quantizer)
@@ -635,7 +643,7 @@ def permute(w, heads):
         if args.ptq != "16a8w":
             # 16a8w use 16bit kv io, so skip this custom annotation
             custom_annotations = custom_annotations + (annotate_matmul_16a8w,)
-        if args.decoder_model in {"stories110m", "stories260k"}:
+        if args.decoder_model in {"stories110m", "stories260k", "phi_4_mini"}:
             custom_annotations = custom_annotations + (
                 annotate_linear_16a8w_in_affine_layer,
             )
@@ -853,12 +861,20 @@ def post_process():
 
     seq_len = args.max_seq_len
     multi_prompts = " ".join([f'--prompt "{prompt}"' for prompt in args.prompt])
+    lookahead_args = " ".join(
+        [
+            f"--window {args.window}",
+            f"--gcap {args.gcap}",
+            f"--ngram {args.ngram}",
+        ]
+    )
     runner_args = " ".join(
         [
             multi_prompts,
             f"--eval_mode {EVAL_MODE[args.model_mode]}",
             f"--temperature {args.temperature}",
             f"--system_prompt '{args.system_prompt}'",
+            lookahead_args if args.model_mode == "lookahead" else "",
         ]
     )
 
@@ -908,9 +924,6 @@ def post_process():
                 "--output_path outputs/outputs.txt",
                 f"--performance_output_path {performance_output_path}",
                 f"--kv_updater {'SmartMask' if args.kv_updater == smart_mask_updater else 'ShiftPointer'}",
-                f"--window {args.window}",
-                f"--gcap {args.gcap}",
-                f"--ngram {args.ngram}",
                 runner_args,
             ]
         )
@@ -1175,6 +1188,13 @@ def _build_parser():
         action="store_true",
         default=False,
     )
+    parser.add_argument(
+        "-G",
+        "--group_size",
+        type=int,
+        default=None,
+        help="group_size used in block quantization for weight quantization.",
+    )
 
     parser.add_argument("-v", "--verbose", action="store_true")