shared buffer + smart mask

- Add flag to use smart mask or shift pointer
- Add llama3_2 python with smart mask updator
- Change Memory class to IoMgrBase
- Change HybridMemory class to ShiftPointerIoMgr

Author: Chun-I Tsai

backends/qualcomm/runtime/QnnManager.cpp

@@ -154,8 +154,9 @@ Error QnnManager::RegisterMem(
     const std::shared_ptr<TensorWrapper>& tensor_wrapper) {
   SharedBuffer& shared_buffer_manager = SharedBuffer::GetSharedBufferManager();
   // Not enable shared buffer
-  if (!options_->shared_buffer())
+  if (!options_->shared_buffer()) {
     return Error::Internal;
+  }
 
   if (backend_params_ptr_->qnn_mem_manager_ptr_ == nullptr) {
     QNN_EXECUTORCH_LOG_WARN(

backends/qualcomm/runtime/QnnManager.h

@@ -145,7 +145,7 @@ class QnnManager {
           {Qnn_DataType_t::QNN_DATATYPE_UFIXED_POINT_8,
            executorch::aten::ScalarType::Byte},
           {Qnn_DataType_t::QNN_DATATYPE_UFIXED_POINT_16,
-           executorch::aten::ScalarType::Bits16},
+           executorch::aten::ScalarType::UInt16},
   };
 };
 } // namespace qnn

backends/qualcomm/runtime/backends/QnnMemManager.h

@@ -77,7 +77,7 @@ class QnnMemManager {
            Qnn_DataType_t::QNN_DATATYPE_SFIXED_POINT_16},
           {executorch::aten::ScalarType::Byte,
            Qnn_DataType_t::QNN_DATATYPE_UFIXED_POINT_8},
-          {executorch::aten::ScalarType::Bits16,
+          {executorch::aten::ScalarType::UInt16,
            Qnn_DataType_t::QNN_DATATYPE_UFIXED_POINT_16},
   };
 };

examples/qualcomm/oss_scripts/llama2/runner/runner.h

@@ -195,6 +195,7 @@ class IoMemMgr {
           {executorch::aten::ScalarType::Short, sizeof(int16_t)},
           {executorch::aten::ScalarType::Byte, sizeof(uint8_t)},
           {executorch::aten::ScalarType::Bits16, sizeof(uint16_t)},
+          {executorch::aten::ScalarType::UInt16, sizeof(uint16_t)},
   };
 };
 

examples/qualcomm/oss_scripts/llama3_2/CMakeLists.txt

@@ -28,8 +28,8 @@ list(
   ${CMAKE_CURRENT_LIST_DIR}/qnn_llama3_2_runner.cpp
   ${CMAKE_CURRENT_LIST_DIR}/runner/runner.cpp
   ${CMAKE_CURRENT_LIST_DIR}/runner/runner.h
-  ${CMAKE_CURRENT_LIST_DIR}/runner/io_memory.cpp
-  ${CMAKE_CURRENT_LIST_DIR}/runner/io_memory.h
+  ${CMAKE_CURRENT_LIST_DIR}/runner/io_manager.cpp
+  ${CMAKE_CURRENT_LIST_DIR}/runner/io_manager.h
 )
 
 list(

examples/qualcomm/oss_scripts/llama3_2/llama.py

@@ -66,12 +66,42 @@
 logging.getLogger().setLevel(logging.INFO)
 
 
+def smart_mask_updator(atten_mask, pos, k_caches, v_caches, new_k_caches, new_v_caches):
+    for i, k_cache in enumerate(k_caches):
+        k_cache[:, :, pos] = new_k_caches[i][:, :, 0]
+
+    for i, v_cache in enumerate(v_caches):
+        v_cache[:, pos, :] = new_v_caches[i]
+
+    atten_mask[0][pos] = 0
+    pos += 1
+    return (atten_mask, pos, k_caches, v_caches)
+
+
+def shift_pointer_updator(
+    atten_mask, pos, k_caches, v_caches, new_k_caches, new_v_caches
+):
+    k_caches = [
+        torch.cat([k_cache[:, :, 1:], new_k_caches[i]], dim=-1)
+        for i, k_cache in enumerate(k_caches)
+    ]
+    v_caches = [
+        torch.cat([v_cache[:, 1:, :], new_v_caches[i]], dim=1)
+        for i, v_cache in enumerate(v_caches)
+    ]
+
+    pos += 1
+    atten_mask[0][-pos - 1] = 0
+    return (atten_mask, pos, k_caches, v_caches)
+
+
 def _kv_calibrate(
     example_inputs,
     user_prompts,
     module: torch.fx.GraphModule,
     tokenizer_model_path="tokenizer.model",
     max_seq_len=512,
+    updator=smart_mask_updator,
 ):
     sp_model = get_tokenizer(tokenizer_model_path)
     _, atten_mask, _, k_caches, v_caches = example_inputs
@@ -92,17 +122,9 @@ def _kv_calibrate(
                 *k_caches,
                 *v_caches,
             )
-            k_caches = [
-                torch.cat([k_cache[:, :, 1:], new_k_caches[i]], dim=-1)
-                for i, k_cache in enumerate(k_caches)
-            ]
-            v_caches = [
-                torch.cat([v_cache[:, 1:, :], new_v_caches[i]], dim=1)
-                for i, v_cache in enumerate(v_caches)
-            ]
-
-            pos += 1
-            atten_mask[0][-pos - 1] = 0
+            atten_mask, pos, k_caches, v_caches = updator(
+                atten_mask, pos, k_caches, v_caches, new_k_caches, new_v_caches
+            )
             if pos >= len(token_list):
                 token_list.append(torch.argmax(logits[:, -1], dim=-1).item())
 
@@ -153,6 +175,7 @@ def calibrate(
     module: torch.fx.GraphModule,
     tokenizer_model_path="tokenizer.model",
     max_seq_len=512,
+    kv_updator=smart_mask_updator,
 ):
     if len(example_inputs) == 2:
         _prefill_calibrate(
@@ -169,6 +192,7 @@ def calibrate(
             module,
             tokenizer_model_path,
             max_seq_len,
+            updator=kv_updator,
         )
     else:
         raise RuntimeError("Get wrong inputs")
@@ -298,13 +322,15 @@ def quantize(self, quant_dtype, args, custom_annotations=()):
                 self.llama_model, self.inputs, strict=True
             ).module()
             fx_graph_module = prepare_pt2e(fx_graph_module, quantizer)
+
         logging.info("Quantizing the model...")
         calibrate(
             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=self.llama_meta["get_max_seq_len"],
+            kv_updator=args.kv_updator,
         )
 
         self.llama_model = convert_pt2e(fx_graph_module)
@@ -316,6 +342,7 @@ def lowering_modules(
         use_fp16=False,
         soc_model=QcomChipset.SM8650,
         num_sharding=0,
+        shared_buffer=False,
     ):
         executorch_config = ExecutorchBackendConfig(
             # For shared buffer, user must pass the memory address
@@ -336,7 +363,7 @@ def lowering_modules(
             compiler_specs = generate_qnn_executorch_compiler_spec(
                 soc_model=soc_model,
                 backend_options=backend_options,
-                shared_buffer=False,
+                shared_buffer=shared_buffer,
             )
             skip_node_op_set = {"llama.fallback.default"}
             partitioner = QnnPartitioner(
@@ -366,7 +393,7 @@ def lowering_modules(
             if num_sharding > 0:
                 update_spill_fill_size(edge_prog_mgr.exported_program())
             exec_prog_mgr = edge_prog_mgr.to_executorch(config=executorch_config)
-            with open(f"{work_space}/{pte_filename}.pte", "wb") as file:
+            with open(f"{work_space}/{self.pte_filename}.pte", "wb") as file:
                 exec_prog_mgr.write_to_file(file)
 
     def get_example_inputs(self, use_kv_cache=True):
@@ -491,6 +518,7 @@ def compile(args, pte_filename):
             use_fp16=use_fp16,
             soc_model=get_soc_to_chipset_map()[args.model],
             num_sharding=args.num_sharding,
+            shared_buffer=args.shared_buffer,
         )
         quant_attrs = llama_instance_list[0].get_quant_attrs()
     else:
@@ -525,7 +553,7 @@ def compile(args, pte_filename):
             generate_qnn_executorch_compiler_spec(
                 soc_model=get_soc_to_chipset_map()[args.model],
                 backend_options=backend_options,
-                shared_buffer=True,
+                shared_buffer=args.shared_buffer,
                 multiple_graphs=True,
                 graph_name=graph_name,
             )
@@ -697,6 +725,7 @@ def inference(args, quant_attrs, pte_filename, pre_gen_pte=""):
             f"--system_prompt '{args.system_prompt}'",
             f"--logits_scale {quant_attrs['scale']}",
             f"--logits_offset {quant_attrs['zero_point']}",
+            f"--kv_updator {'SmartMask' if args.kv_updator == smart_mask_updator else 'ShiftPointer'}",
         ]
     )
     runner_cmd = " ".join(
@@ -862,6 +891,14 @@ def main():
         type=int,
     )
 
+    parser.add_argument(
+        "--kv_updator",
+        help="Choose how to update kv cache during runtime",
+        choices=["smart_mask", "shift_pointer"],
+        default="smart_mask",
+        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")
@@ -878,6 +915,14 @@ def main():
     else:
         raise RuntimeError(f"No such model_mode {args.model_mode}.")
 
+    if args.kv_updator == "smart_mask":
+        args.shared_buffer = True
+        args.kv_updator = smart_mask_updator
+    elif args.kv_updator == "shift_pointer":
+        args.kv_updator = shift_pointer_updator
+    else:
+        exit(f"Using an unkown kv update {args.kv_updator}")
+
     if args.pre_gen_pte:
         quant_attrs = json.load(
             open(f"{args.pre_gen_pte}/{pte_filename}_quant_attrs.txt")

examples/qualcomm/oss_scripts/llama3_2/qnn_llama3_2_runner.cpp

@@ -50,6 +50,10 @@ DEFINE_int32(
     "0: PromptProcessor(prefill) / 1: TokenGenerator(kv) / 2: HybridMode (prefill+kv)");
 DEFINE_double(logits_scale, 0.0, "Logits scale");
 DEFINE_int32(logits_offset, 0, "Logits offset");
+DEFINE_string(
+    kv_updator,
+    "How to update kv cache. Choose between SmartMask and ShiftPointer",
+    "SmartMask");
 
 int main(int argc, char** argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
@@ -61,7 +65,8 @@ int main(int argc, char** argv) {
       FLAGS_logits_scale,
       FLAGS_logits_offset,
       FLAGS_temperature,
-      FLAGS_eval_mode);
+      FLAGS_eval_mode,
+      FLAGS_kv_updator);
   std::vector<char> buf;
   buf.reserve(5 * FLAGS_seq_len); // assume each token is around 5 char
   std::ofstream fout(FLAGS_output_path.c_str());