allocate workspace from allocator for kunlun runtime

.gitmodules

@@ -13,6 +13,3 @@
 [submodule "example"]
 	path = examples/NNmodel
 	url = git@github.com:wanghailu0717/NNmodel.git
-[submodule "examples/distributed/onnxsim_large_model"]
-	path = examples/distributed/onnxsim_large_model
-	url = git@github.com:luchangli03/onnxsim_large_model.git

Makefile

@@ -40,7 +40,7 @@ endif
 
 build:
 	mkdir -p build/$(TYPE)
-	cd build/$(TYPE) && cmake $(CMAKE_OPT) ../.. && make -j8
+	cd build/$(TYPE) && cmake $(CMAKE_OPT) ../.. && make -j64
 
 clean:
 	rm -rf build

examples/distributed/README.md

@@ -1,7 +1,5 @@
 # 分布式脚本
 
-## 英伟达平台运行方式
-
 #### 1. 运行pytorch模型并生成输入和标准输出，可选择导出onnx
 
 使用 `--export_onnx` 设置导出onnx的目录，默认为当前路径 `./`，不使用这个flag则只进行计算和生成输入输出。
@@ -17,23 +15,3 @@ python run_pytorch.py --model gpt2  --batch_size 1  --length 1 --export_onnx ./
 ```bash
 python cuda_launch.py --model "/XXX/XXX.onnx" --nproc_per_node 4 
 ```
-
-## 寒武纪平台运行方式
-
-**将上述运行脚本 `run_pytorch.py` 以及 `cuda_launch.py` 针对寒武纪平台做了相应的适配，具体见 `run_pytorch_mlu.py` 以及 `bang_launch.py`。**
-
-#### 1. 运行pytorch模型并生成输入和标准输出，可选择导出onnx
-
-使用 `--export_onnx` 设置导出onnx的目录，默认为当前路径 `./`，不使用这个flag则只进行计算和生成输入输出。
-
-```bash
-python run_pytorch_mlu.py --model gpt2  --batch_size 1  --length 1 --export_onnx ./
-```
-
-会在当前目录下生成输入输出文件`test_inputs.npy` 和 `test_results.npy`，目前只支持单一输入输出。
-
-#### 2. 运行InfiniTensor分布式脚本
-
-```bash
-python bang_launch.py --model "/XXX/XXX.onnx" --nproc_per_node 4 
-```

examples/distributed/bang/bang_launch.py

@@ -1,39 +1,35 @@
-import sys
-sys.path.append('../')
-
 import argparse
 import os
 import time
 import multiprocessing as mp
 from pyinfinitensor.onnx import OnnxStub, backend
 import onnx
-from onnx.external_data_helper import convert_model_to_external_data
 from onnx.shape_inference import infer_shapes_path
 import numpy as np
 from parallel_opt import parallel_model
 
+
 def parse_args():
     parser = argparse.ArgumentParser(description="launch distributed infinitensor")
     parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
     parser.add_argument(
-        "--nproc_per_node", type=int, default=1, help="number of processes per node"
+        "--nproc_per_node", type=int, default=2, help="number of processes per node"
     )
     parser.add_argument(
         "--name", type=str, default="test", help="name of this instance."
     )
     parser.add_argument(
-        "--model", type=str, required=True, help="path to the ONNX model file."
+        "--model", type=str, default="/data/onnx_models/llama2/llama_bs1_seq1024.onnx", 
+        help="path to the ONNX model file."
     )
     parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
     parser.add_argument("--length", type=int, default=1, help="sequence length.")
     parser.add_argument(
         "--gen_std",
+        default=False,
         action="store_true",
         help="whether to generate the standard results.",
     )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], default="fp32", help="data type"
-    )
     args = parser.parse_args()
     print("arg setting: ", args)
     return (
@@ -44,46 +40,39 @@ def parse_args():
         args.batch_size,
         args.length,
         args.gen_std,
-        args.type,
     )
 
 
-def run_model(model, runtime, world_size=1, rank=0, n=10, data_type="default"):
-    stub = OnnxStub(model, runtime, matmul_compute_type=data_type)
+def run_model(model, runtime, world_size=1, rank=0, n=10):
+    stub = OnnxStub(model, runtime)
     load_inputs(stub, world_size, rank)
     # stub.tune()
     stub.run()
     # get outputs
+    time.sleep(0.01)
     outputs = next(stub.outputs.values().__iter__()).copyout_numpy()
 
     # bench
+    begin = time.time()
     for _ in range(n):
         stub.run()
-    begin = time.time()
-    for _ in range(n * 2):
-        stub.run()
     end = time.time()
-    avg_time = (end - begin) / (n * 2)
+    avg_time = (end - begin) / n
     print(f"average time: {avg_time}")
     return outputs
 
-def load_inputs(stub, world_size=1, rank=0):
-    for i, (name, tensor) in enumerate(stub.inputs.items()):
-        input = np.load(f"./data/input_{i}.npy")
-        if all(x == y for x,y in zip(input.shape,tensor.shape())):
-            tensor.copyin_numpy(input)
-        else:
-            tensor.copyin_numpy(np.hsplit(input, world_size)[rank])
 
-
-def run_and_compare(name, model, runtime, world_size=1, rank=0, data_type="default"):
+def run_and_compare(name, model, runtime, world_size=1, rank = 0):
     results = np.load(f"./data/output.npy")
-    outputs = run_model(model, runtime, world_size, rank, data_type=data_type)
-    print("outputs abs mean:", abs(outputs).mean())
-    print("max abs diff:", abs(outputs - results).max())
+    outputs = run_model(model, runtime, world_size, rank)
+    print("answer argmax:", np.argmax(results))
+    print("output argmax:", np.argmax(outputs))
+    #np.testing.assert_allclose(outputs, results, rtol=1e-3, atol=1e-3)
+    getDiff(results, outputs)
+
 
 def start_worker(
-    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto, data_type: str
+    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto
 ):
     dist_name = name + "_dist"
     model = parallel_model(model, world_size, rank)
@@ -96,7 +85,7 @@ def start_worker(
         save_as_external_data=True,
         location=extern_path,
     )
-    #infer_shapes_path(f"./{dist_name}_rank{rank}.onnx")
+    infer_shapes_path(f"./{dist_name}_rank{rank}.onnx")
     runtime = backend.BangRuntime(local_rank)
     # print("init comm")
     runtime.init_comm(
@@ -104,12 +93,13 @@ def start_worker(
         world_size,
         rank,
     )
-    run_and_compare(name, model, runtime, world_size, rank, data_type)
+    run_and_compare(name, model, runtime, world_size, rank)
 
 
-def start_single(name, model, data_type):
+def start_single(name, model):
     runtime = backend.BangRuntime(0)
-    run_and_compare(name, model, runtime, data_type=data_type)
+    run_and_compare(name, model, runtime)
+
 
 def generate_input_output(model):
     os.makedirs(os.path.dirname("./data/"), exist_ok=True)
@@ -142,36 +132,55 @@ def generate_input_output(model):
     np.save(f"./data/output", output)
 
 
+def load_inputs(stub, world_size=1, rank=0):
+    for i, (name, tensor) in enumerate(stub.inputs.items()):
+        input = np.load(f"./data/input_{i}.npy")
+        if all(x == y for x,y in zip(input.shape,tensor.shape())):
+            tensor.copyin_numpy(input)
+        else:
+            tensor.copyin_numpy(np.hsplit(input, world_size)[rank])
+
+def getDiff(base, test):
+    absolute_diff = np.abs(np.subtract(base, test))
+    max_absolute_diff = np.max(absolute_diff)
+
+    baseCopy = base.astype(np.float64).ravel()
+    testCopy = test.astype(np.float64).ravel()
+    upValue = np.sum(np.abs(baseCopy - testCopy))
+    downValue = np.sum(np.abs(baseCopy)) + np.float64(1e-9)
+    max_relative_diff = upValue / downValue
+    print(f"Max absolute difference: {max_absolute_diff}\n"
+          f"Max relative difference: {max_relative_diff}")
+    return max_absolute_diff, max_relative_diff
+
+
 def main():
-    nnodes, nproc_per_node, name, model_path, bs, length, gen_std, data_type = parse_args()
-    data_type = "default" if data_type == "fp32" else data_type
-    
+    nnodes, nproc_per_node, name, model_path, bs, length, gen_std = parse_args()
+
     model = onnx.load(model_path)
 
     # generate standart output
     if gen_std:
-        print(f"generate standard data for {name}.")
-        # a small vocabulary size to fit all LLM.
-        generate_input_output(model)
+        print("Generate inputs and outputs.")
+        p = mp.Process(target=generate_input_output, args=[model])
+        p.start()
+        p.join()
         return
 
-    if nproc_per_node == 1:
-        # run single process.
-        # use standalone process to isolate bang.
-        print("run model by single MLU.")
-        # p = mp.Process(target=start_single, args=(name, model, data_type))
-        # p.start()
-        # p.join()
-        start_single(name, model, data_type)
-        return
+    # run single process.
+    # use standalone process to isolate cuda.
+    print("run model by single MLU.")
+    p = mp.Process(target=start_single, args=(name, model))
+    p.start()
+    p.join()
 
     # run distributed parallel.
     world_size = nnodes * nproc_per_node
-    print(f"run model by {world_size} MLU in parallel.")
+    print(f"run model by {world_size} MLUs in parallel.")
     workers = [
         mp.Process(
             target=start_worker,
-            args=(name, world_size, rank, rank % nproc_per_node, model, data_type),
+            args=(name, world_size, rank, rank % nproc_per_node, model),
         )
         for rank in range(world_size)
     ]

examples/distributed/bang/run_pytorch_mlu.py

@@ -1,249 +0,0 @@
-import argparse
-import torch
-import torch_mlu
-from transformers import BertModel, BertConfig
-from transformers import GPT2Model, GPT2Config
-from transformers import OPTModel, OPTConfig
-from transformers import AlbertModel, AlbertConfig
-from transformers import LlamaModel, LlamaConfig
-import time
-import numpy as np
-import onnx
-import sys
-import os
-from onnx.external_data_helper import convert_model_to_external_data
-from onnxsim import simplify
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Run pytorch gpt2/bert/opt and optionally export onnx.")
-    parser.add_argument(
-        "--model", type=str, choices=["gpt2", "bert", "opt", "llama", "albert"], required=True, help="model type"
-    )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--export_onnx",
-        type=str,
-        nargs="?",
-        default=None,
-        const="./",
-        help="whether and where to export onnx file",
-    )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], required=True, help="model data type"
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.model,
-        args.batch_size,
-        args.length,
-        args.export_onnx,
-        args.type
-    )
-
-
-def get_model(modelname):
-    match modelname:
-        case "albert":
-            model = AlbertModel.from_pretrained("albert/albert-base-v2")
-            voc_size = AlbertConfig().vocab_size
-        case "bert":
-            model = BertModel.from_pretrained("bert-base-uncased", add_pooling_layer=False, hidden_act="gelu_new") # erf is not impl by infini
-            voc_size = BertConfig().vocab_size
-        case "gpt2":
-            model = GPT2Model.from_pretrained("GPT2")
-            voc_size = GPT2Config().vocab_size
-        case "opt":
-            model = OPTModel.from_pretrained("facebook/opt-125m")
-            voc_size = OPTConfig().vocab_size
-        case "llama":
-            model = LlamaModel.from_pretrained("meta-llama/Llama-2-7b-hf")
-            voc_size = LlamaConfig().vocab_size
-        case _:
-            raise KeyError(modelname)
-
-    model = model.eval()
-    return model, voc_size
-
-def run_pytorch(torch_model, voc_size, batchsize, len, dtype="fp32"):
-    data = np.random.randint(0, voc_size, (batchsize, len), dtype=np.int32)
-    os.makedirs(os.path.dirname("./data/"), exist_ok=True)
-    np.save("./data/input_0", data)
-    inputs = torch.from_numpy(data).to("mlu")
-    torch_model = torch_model.to("mlu")
-    if dtype == "fp16":
-        torch_model = torch_model.half()
-
-    n_iter = 20
-    with torch.no_grad():
-        for _ in range(10):
-            outputs = torch_model(inputs)
-    torch.mlu.synchronize()
-    begin = time.time()
-    with torch.no_grad():
-        for _ in range(n_iter):
-            torch.mlu.synchronize()
-            outputs = torch_model(inputs)
-            torch.mlu.synchronize()
-    torch.mlu.synchronize()
-    end = time.time()
-    
-    avg_time = (end - begin) / n_iter
-    outputs = outputs.last_hidden_state.to("cpu")
-    print("outputs abs mean:", abs(np.array(outputs)).mean())
-    print(f"average time: {avg_time}")
-    # torch.mlu.memory.empty_cache()
-    np.save("./data/output", np.array(outputs))
-    print("Save input & output into ./data.")
-
-
-def export_onnx(modelname, model, data, path, extern=False, dtype="fp32"):
-    data = data.to("mlu")
-    model = model.to("mlu")
-    if dtype == "fp16":
-        model = model.half()
-    torch.onnx.export(model, data, path, verbose=False, do_constant_folding=True)
-    if modelname != "llama":
-        # use onnxsim to simplify
-        onnx_model = onnx.load(path)
-        onnx_model, check = simplify(onnx_model, skipped_optimizers=['eliminate_duplicate_initializer'])
-        # onnx_model, check = simplify(onnx_model, skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
-        assert check
-        add_value_info_for_constants(onnx_model)
-        onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
-        if extern:
-            extern_path = path.replace('.onnx', '.pb')
-            if os.path.exists(extern_path):
-                os.remove(extern_path)
-            extern_path = extern_path.split("/")[-1]
-            convert_model_to_external_data(
-                onnx_model,
-                all_tensors_to_one_file=True,
-                location=extern_path,
-                size_threshold=1024,
-                convert_attribute=False,
-            )
-        onnx.save(onnx_model, path)
-    else:
-        # use third party tool to simplify llama
-        # reference: https://github.com/luchangli03/onnxsim_large_model/
-        sys.path.append("onnxsim_large_model")
-        from onnx_utils import set_onnx_input_shape
-        from compress_model import SIZE_1MB, compress_onnx_model, uncompress_onnx_model
-
-        in_model_path = path
-        out_model_path = path
-        if not out_model_path:
-            out_model_path = in_model_path[:-5] + ".sim.onnx"
-        if os.path.isdir(out_model_path):
-            out_model_path = os.path.join(out_model_path, os.path.basename(in_model_path))
-
-        onnx_model = onnx.load(in_model_path)
-        print(f"load model from {in_model_path} success")
-
-        size_th_bytes = 1024 * 1024
-
-        onnx_model, removed_inits = compress_onnx_model(onnx_model, size_th_bytes=size_th_bytes)
-        print(f"compress model success")
-
-        onnx_model = set_onnx_input_shape(onnx_model, "")
-
-        tensor_size_threshold = f"1024KB"
-        skipped_optimizers = []
-        skipped_optimizers.append("eliminate_duplicate_initializer")
-        onnx_model, check = simplify(onnx_model, skipped_optimizers=skipped_optimizers,
-                                    tensor_size_threshold=tensor_size_threshold)
-        if not check:
-            raise ValueError(f"simplify compressed model {in_model_path} failed")
-
-        print(f"simplify model success")
-
-        onnx_model = uncompress_onnx_model(onnx_model, removed_inits)
-        print(f"uncompress model success")
-
-        add_value_info_for_constants(onnx_model)
-
-        onnx.save(onnx_model, out_model_path, save_as_external_data=True)
-
-
-def add_value_info_for_constants(model : onnx.ModelProto):
-    """
-    Currently onnx.shape_inference doesn't use the shape of initializers, so add
-    that info explicitly as ValueInfoProtos.
-    Mutates the model.
-    Args:
-        model: The ModelProto to update.
-    """
-    # All (top-level) constants will have ValueInfos before IRv4 as they are all inputs
-    if model.ir_version < 4:
-        return
-
-    def add_const_value_infos_to_graph(graph : onnx.GraphProto):
-        inputs = {i.name for i in graph.input}
-        existing_info = {vi.name: vi for vi in graph.value_info}
-        for init in graph.initializer:
-            # Check it really is a constant, not an input
-            if init.name in inputs:
-                continue
-
-            # The details we want to add
-            elem_type = init.data_type
-            shape = init.dims
-
-            # Get existing or create new value info for this constant
-            vi = existing_info.get(init.name)
-            if vi is None:
-                vi = graph.value_info.add()
-                vi.name = init.name
-
-            # Even though it would be weird, we will not overwrite info even if it doesn't match
-            tt = vi.type.tensor_type
-            if tt.elem_type == onnx.TensorProto.UNDEFINED:
-                tt.elem_type = elem_type
-            if not tt.HasField("shape"):
-                # Ensure we set an empty list if the const is scalar (zero dims)
-                tt.shape.dim.extend([])
-                for dim in shape:
-                    tt.shape.dim.add().dim_value = dim
-
-        # Handle subgraphs
-        for node in graph.node:
-            for attr in node.attribute:
-                # Ref attrs refer to other attrs, so we don't need to do anything
-                if attr.ref_attr_name != "":
-                    continue
-
-                if attr.type == onnx.AttributeProto.GRAPH:
-                    add_const_value_infos_to_graph(attr.g)
-                if attr.type == onnx.AttributeProto.GRAPHS:
-                    for g in attr.graphs:
-                        add_const_value_infos_to_graph(g)
-
-
-    return add_const_value_infos_to_graph(model.graph)
-
-
-def main():
-    torch.backends.mlu.matmul.allow_tf32 = False
-    torch.backends.cnnl.allow_tf32 = False
-    modelname, batchsize, seqlen, export_path, dtype = parse_args()
-    if dtype == "tf32":
-        torch.backends.mlu.matmul.allow_tf32 = True
-    else:
-        os.environ["CAMBRICON_TF32_OVERRIDE"] = "0"
-
-    model, voc_size = get_model(modelname)
-    if export_path is not None:
-        filename = "{}_{}_{}_{}.onnx".format(modelname, batchsize, seqlen, dtype)
-        path = os.path.join(export_path, filename)
-        if not os.path.exists(path):
-            param = torch.zeros((batchsize, seqlen), dtype=torch.int)
-            export_onnx(modelname, model, param, path, True, dtype)
-        else:
-            print("Onnx path exists, skipping export.")
-
-    run_pytorch(model, voc_size, batchsize, seqlen, dtype)
-
-if __name__ == "__main__":
-    main()

examples/distributed/cuda/cuda_launch.py

@@ -10,6 +10,9 @@ import numpy as np
 from parallel_opt import parallel_model
 
 
+os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
+
+
 def parse_args():
     parser = argparse.ArgumentParser(description="launch distributed infinitensor")
     parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
@@ -29,9 +32,6 @@ def parse_args():
         action="store_true",
         help="whether to generate the standard results.",
     )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], default="fp32", help="data type"
-    )
     args = parser.parse_args()
     print("arg setting: ", args)
     return (
@@ -42,12 +42,11 @@ def parse_args():
         args.batch_size,
         args.length,
         args.gen_std,
-        args.type,
     )
 
 
-def run_model(model, runtime, inputs, n=10, data_type = "default"):
-    stub = OnnxStub(model, runtime, matmul_compute_type=data_type)
+def run_model(model, runtime, inputs, n=10):
+    stub = OnnxStub(model, runtime)
     for tensor, input in zip(stub.inputs.values(), inputs, strict=False):
         tensor.copyin_numpy(input)
     # stub.tune()
@@ -67,17 +66,17 @@ def run_model(model, runtime, inputs, n=10, data_type = "default"):
     return outputs
 
 
-def run_and_compare(name, model, runtime, data_type):
+def run_and_compare(name, model, runtime):
     input_ids = np.load(f"{name}_inputs.npy")
     position_ids = np.arange(input_ids.shape[-1])
     results = np.load(f"{name}_results.npy")
-    outputs = run_model(model, runtime, (input_ids, position_ids), data_type=data_type)
+    outputs = run_model(model, runtime, (input_ids, position_ids))
     print("outputs abs mean:", abs(outputs).mean())
     print("max abs diff:", abs(outputs - results).max())
 
 
 def start_worker(
-    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto, data_type: str
+    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto
 ):
     dist_name = name + "_dist"
     model = parallel_model(model, world_size, rank)
@@ -98,12 +97,12 @@ def start_worker(
         world_size,
         rank,
     )
-    run_and_compare(name, model, runtime, data_type)
+    run_and_compare(name, model, runtime)
 
 
-def start_single(name, model, data_type):
+def start_single(name, model):
     runtime = backend.CudaRuntime(0)
-    run_and_compare(name, model, runtime, data_type)
+    run_and_compare(name, model, runtime)
 
 
 def gen_standard(name, model, voc_size, bs, len):
@@ -118,10 +117,8 @@ def gen_standard(name, model, voc_size, bs, len):
 
 
 def main():
-    nnodes, nproc_per_node, name, model_path, bs, length, gen_std, data_type = parse_args()
-    data_type = "default" if data_type == "fp32" else data_type
-    if data_type != "tf32":
-        os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
+    nnodes, nproc_per_node, name, model_path, bs, length, gen_std = parse_args()
+
     model = onnx.load(model_path)
 
     # generate standart output
@@ -135,7 +132,7 @@ def main():
     # run single process.
     # use standalone process to isolate cuda.
     print("run model by single GPU.")
-    p = mp.Process(target=start_single, args=(name, model, data_type))
+    p = mp.Process(target=start_single, args=(name, model))
     p.start()
     p.join()
 
@@ -145,7 +142,7 @@ def main():
     workers = [
         mp.Process(
             target=start_worker,
-            args=(name, world_size, rank, rank % nproc_per_node, model, data_type),
+            args=(name, world_size, rank, rank % nproc_per_node, model),
         )
         for rank in range(world_size)
     ]

examples/distributed/kunlun/export_onnx.sh

@@ -1,14 +0,0 @@
- export HF_ENDPOINT=https://hf-mirror.com
-
-models=("bert" "gpt2" "llama")
-batch_size=(1 32)
-seq_len=(100 500)
-nproc=(1 2 4)
-
-for model in "${models[@]}"; do
-    for bs in "${batch_size[@]}"; do
-        for len in "${seq_len[@]}"; do
-            python run_pytorch.py --model "$model" --batch_size "$bs" --length "$len" --export_onnx ../models/"$model" --export_only 
-        done
-    done
-done 

examples/distributed/kunlun/kunlun_launch.py

@@ -21,7 +21,7 @@ def parse_args():
         "--nproc_per_node", type=int, default=2, help="number of processes per node"
     )
     parser.add_argument(
-        "--name", type=str, choices=["gpt2", "bert", "llama"], help="name of model."
+        "--name", type=str, default="test", help="name of this instance."
     )
     parser.add_argument(
         "--model", type=str, default="", help="path to the ONNX model file."
@@ -54,12 +54,12 @@ def parse_args():
         help="path to save internal onnx model for parallel run"
     )
     args = parser.parse_args()
-
+    
     # check path, mkdir if not exist
     check_exists(args.input_dir)
     check_exists(args.result_dir)
     check_exists(args.internal_model_dir)
-
+    
     print("arg setting: ", args)
     return (
         args.num_nodes,
@@ -97,7 +97,7 @@ def perf_it(n):
             # warmup
             for _ in range(n):
                 func(*args, **kwargs)
-
+            
             t_total = 0
             for _ in range(n):
                 t0 = time.time()
@@ -112,22 +112,20 @@ def perf_it(n):
 
 
 """
-Run InfiniTensor model with Standard input
+Run InfiniTensor model with Standard input 
 check=True: check with standard output gen by pytorch
 perf=True: run n times to get avg time
 """
 def run_model(task_name,
-              model,
-              runtime,
-              world_size=1,
-              rank=0,
+              model, 
+              runtime, 
+              world_size=1, 
+              rank=0, 
               n=10,
               check=True,
               perf=True):
-
-    stub = OnnxStub(model, runtime,
-                    use_naive_allocator=True \
-                    if task_name == "llama" else False)
+    
+    stub = OnnxStub(model, runtime)
 
     # load in Onnx model inputs
     def load_inputs(stub: OnnxStub):
@@ -163,9 +161,9 @@ def run_model(task_name,
         if np.isnan(output).any():
             print("Nan in output")
             exit()
-        np_assert(st_output, output)
+        np_assert(st_output, output)        
 
-    # perf
+    # perf 
     if perf:
         @perf_it(n)
         def perf_infinitensor(stub: OnnxStub):
@@ -178,17 +176,17 @@ def run_model(task_name,
 """
 Start a worker in Parallel
 """
-def start_worker(name: str,
-           world_size: int,
-           rank: int,
-           local_rank: int,
+def start_worker(name: str, 
+           world_size: int, 
+           rank: int, 
+           local_rank: int, 
            model: onnx.ModelProto):
 
     dist_name = name + "_dist"
     # partial a onnx model to world_size part
     model = parallel_model(model, world_size, rank)
     onnx.save(model, os.path.join(internal_model_dir, \
-                                    f"{dist_name}_rank{rank}.onnx"), save_as_external_data=True)
+                                    f"{dist_name}_rank{rank}.onnx"))
     runtime = backend.KUNLUNRuntime(local_rank)
     # print("init comm")
     runtime.init_comm(
@@ -200,10 +198,10 @@ def start_worker(name: str,
 
 
 """
-generate standard input/output with
+generate standard input/output with 
 sigle card run
 """
-def gen_standard(task_name: str, model: onnx.ModelProto):
+def gen_stardard(task_name: str, model: onnx.ModelProto):
     runtime = backend.KUNLUNRuntime(0)
     stub = OnnxStub(model, runtime)
     position_id = 0
@@ -238,25 +236,25 @@ def gen_standard(task_name: str, model: onnx.ModelProto):
 def main():
 
     global input_dir, result_dir, internal_model_dir
-
+    
     nnodes, nproc_per_node, task_name, \
         model_path, gen_std, run_single, \
             input_dir, result_dir, internal_model_dir = parse_args()
-
+    
     # load input onnx model
     model = onnx.load(model_path)
 
     # generate standart output
     if gen_std:
         print("Generate inputs and outputs.")
-        gen_standard(task_name, model)
+        gen_stardard(task_name, model)
         return
 
     if run_single:
         print("Run model by one GPU card.")
         runtime = backend.KUNLUNRuntime(0)
         run_model(task_name, model, runtime)
-        return
+        return 
 
     # run distributed parallel.
     world_size = nnodes * nproc_per_node

examples/distributed/kunlun/kunlun_launch2.py

@@ -0,0 +1,215 @@
+import sys
+sys.path.append("../")
+import argparse
+import os
+import time
+import multiprocessing as mp
+from pyinfinitensor.onnx import OnnxStub, backend
+import onnx
+from onnx.external_data_helper import convert_model_to_external_data
+from onnx.shape_inference import infer_shapes_path
+import numpy as np
+from parallel_opt import parallel_model
+
+st_input_dir = ".cache/input/"
+st_output_dir = ".cache/output/"
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="launch distributed infinitensor")
+    parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
+    parser.add_argument(
+        "--nproc_per_node", type=int, default=2, help="number of processes per node"
+    )
+    parser.add_argument(
+        "--name", type=str, default="test", help="name of this instance."
+    )
+    parser.add_argument(
+        "--model", type=str, default="/data1/shared/panzezhong/llama/fp32/my_llama_fp32.sim.onnx", help="path to the ONNX model file."
+    )
+    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
+    parser.add_argument("--length", type=int, default=1, help="sequence length.")
+    parser.add_argument(
+        "--gen_std",
+        default=False,
+        action="store_true",
+        help="whether to generate the standard results.",
+    )
+    parser.add_argument(
+        "--run_single",
+        default=False,
+        action="store_true",
+        help="whether run model with single process with standard inputs"
+    )
+    args = parser.parse_args()
+    print("arg setting: ", args)
+    return (
+        args.num_nodes,
+        args.nproc_per_node,
+        args.name,
+        args.model,
+        args.batch_size,
+        args.length,
+        args.gen_std,
+        args.run_single
+    )
+
+
+def run_model(model, runtime, world_size=1, rank=0, n=10):
+    stub = OnnxStub(model, runtime)
+    load_inputs(stub, world_size, rank)
+    # stub.tune()
+    stub.run()
+    # get outputs
+    time.sleep(0.01)
+    outputs = next(stub.outputs.values().__iter__()).copyout_numpy()
+
+    # bench
+    begin = time.time()
+    for _ in range(n):
+        stub.run()
+    end = time.time()
+    avg_time = (end - begin) / n
+    print(f"average time: {avg_time}")
+    return outputs
+
+
+
+def run_and_compare(name, model, runtime, world_size=1, rank = 0):
+    results = np.load(os.path.join(st_output_dir, "test_output.npy"))
+    outputs = run_model(model, runtime, world_size, rank)
+    print(outputs[:100])
+    if np.isnan(outputs).any():
+        print("Nan in output")
+    print("answer argmax:", np.argmax(results))
+    print("output argmax:", np.argmax(outputs))
+    #np.testing.assert_allclose(outputs, results, rtol=1e-3, atol=1e-3)
+    getDiff(results, outputs)
+
+
+def start_worker(
+    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto
+):
+    dist_name = name + "_dist"
+    model = parallel_model(model, world_size, rank)
+    extern_path = f"./{dist_name}_rank{rank}.pb"
+    if os.path.exists(extern_path):
+        os.remove(extern_path)
+    onnx.save_model(
+        model,
+        f"./{dist_name}_rank{rank}.onnx",
+        save_as_external_data=True,
+        location=extern_path,
+    )
+    infer_shapes_path(f"./{dist_name}_rank{rank}.onnx")
+    runtime = backend.KUNLUNRuntime(local_rank)
+    # print("init comm")
+    runtime.init_comm(
+        dist_name,
+        world_size,
+        rank,
+    )
+    run_and_compare(name, model, runtime, world_size, rank)
+
+
+def start_single(name, model):
+    runtime = backend.KUNLUNRuntime(0)
+    run_and_compare(name, model, runtime)
+
+
+def generate_input_output(model):
+    runtime = backend.KUNLUNRuntime(0)
+    stub = OnnxStub(model, runtime)
+    position_id = 0
+    for i, (name, tensor) in enumerate(stub.inputs.items()):
+        input = tensor.copyout_numpy()
+        if np.issubdtype(input.dtype, np.integer):
+            if input.size == 1:
+                # input = np.array([position_id])
+                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
+            else:
+                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
+        elif input.dtype == np.bool_:
+            input = np.random.randint(0,2,size=input.shape) > 0
+        else:
+            if i == 0:
+                input = np.ones(input.shape).astype(input.dtype)
+                position_id = input.shape[-1] - 1
+            else:
+                input = np.random.rand(*input.shape).astype(input.dtype)
+        tensor.copyin_numpy(input)
+        np.save(os.path.join(st_input_dir, f"input_{i}"), input)
+    stub.run()
+    # print(stub.outputs)
+    time.sleep(0.01)
+    output = next(stub.outputs.values().__iter__()).copyout_numpy()
+    print(output[:100])
+    if np.isnan(output).any():
+        print("Nan in output")
+    np.save(os.path.join(st_output_dir, f"output"), output)
+
+
+def load_inputs(stub, world_size=1, rank=0):
+    for i, (name, tensor) in enumerate(stub.inputs.items()):
+        input = np.load(os.path.join(st_input_dir, f"test_input_{name}.npy"))
+        if all(x == y for x,y in zip(input.shape,tensor.shape())):
+            tensor.copyin_numpy(input)
+        else:
+            tensor.copyin_numpy(np.hsplit(input, world_size)[rank])
+
+
+def getDiff(base, test):
+    absolute_diff = np.abs(np.subtract(base, test))
+    max_absolute_diff = np.max(absolute_diff)
+
+    baseCopy = base.astype(np.float64).ravel()
+    testCopy = test.astype(np.float64).ravel()
+    upValue = np.sum(np.abs(baseCopy - testCopy))
+    downValue = np.sum(np.abs(baseCopy)) + np.float64(1e-9)
+    max_relative_diff = upValue / downValue
+    print(f"Max absolute difference: {max_absolute_diff}\nMax relative difference: {max_relative_diff}")
+
+    return max_absolute_diff, max_relative_diff
+
+
+def main():
+    nnodes, nproc_per_node, name, model_path, bs, length, gen_std, run_single = parse_args()
+
+    model = onnx.load(model_path)
+
+    # generate standart output
+    if gen_std:
+        print("Generate inputs and outputs.")
+        p = mp.Process(target=generate_input_output, args=[model])
+        p.start()
+        p.join()
+        return
+
+    # # run single process.
+    # # use standalone process to isolate cuda.
+    if run_single:
+        print("run model by single GPU.")
+        p = mp.Process(target=start_single, args=(name, model))
+        p.start()
+        p.join()
+        return 
+
+    # run distributed parallel.
+    world_size = nnodes * nproc_per_node
+    print(f"run model by {world_size} GPU in parallel.")
+    workers = [
+        mp.Process(
+            target=start_worker,
+            args=(name, world_size, rank, rank % nproc_per_node, model),
+        )
+        for rank in range(world_size)
+    ]
+
+    for w in workers:
+        w.start()
+
+    for w in workers:
+        w.join()
+
+
+if __name__ == "__main__":
+    main()

examples/distributed/kunlun/launch.sh

@@ -1,36 +0,0 @@
-export HF_ENDPOINT=https://hf-mirror.com
-
-# models=("bert" "gpt2" "llama")
-models=("bert" "gpt2")
-batch_size=(1 32)
-seq_len=(100 500)
-nproc=(1 2 4)
-
-results_dir="results"
-
-if [ -d "$results_dir" ]; then
-    echo "directory ./$results_dir exists"
-else
-    mkdir -p "$results_dir"
-    echo "mkdir $results_dir, logs saved there"
-fi
-
-
-for model in "${models[@]}"; do
-    for bs in "${batch_size[@]}"; do
-        for len in "${seq_len[@]}"; do
-            # run pytorch model
-            echo "Run pytorch $model with batch_size=$bs length=$len ."
-            python run_pytorch.py --model "$model" --batch_size "$bs" --length "$len" #> results/"$model"_"$bs"_"$len"_pytorch
-            for n in "${nproc[@]}"; do
-                # run infinitensor 
-                echo "Run $n parallel infinitensor "$model" with batch_size=$bs and length=$len ."
-                python kunlun_launch.py --name "$model" --model ../models/"$model"/"$model"_"$bs"_"$len".onnx --nproc_per_node=$n # >> results/"$model"_"$bs"_"$len"_infini 
-                # delete internal files
-                find ./ -type f -name "*.onnx" -delete
-                find ./ -type f -name "*.pb" -delete
-            done
-            find ./ -type f -name "*.npy" -delete
-        done
-    done
-done

examples/distributed/kunlun/llama_launch.sh

@@ -1,35 +0,0 @@
-export HF_ENDPOINT=https://hf-mirror.com
-
-# models=("bert" "gpt2" "llama")
-models=("llama")
-batch_size=(1 )
-seq_len=(100 500)
-nproc=(1 2 4)
-
-results_dir="results"
-
-if [ -d "$results_dir" ]; then
-    echo "directory ./$results_dir exists"
-else
-    mkdir -p "$results_dir"
-    echo "mkdir $results_dir, logs saved there"
-fi
-
-
-for model in "${models[@]}"; do
-    for bs in "${batch_size[@]}"; do
-        for len in "${seq_len[@]}"; do
-            echo "Run pytorch llama with batch_size="$bs" and length="$len""
-            python run_pytorch.py --model "$model" --batch_size "$bs" --length "$len"
-            for n in "${nproc[@]}"; do
-                    # run pytorch model
-                    echo "Run infinitensor llama with batch_size="$bs" and length="$len" and nproc="$n"."
-                    python kunlun_launch.py --name llama --model ../models/llama/llama_"$bs"_"$len"_fp32.onnx --nproc_per_node=$n
-                    # delete internal files
-                    find ./ -type f -name "*.onnx" -delete
-                    find ./ -type f -name "*0c" -delete
-            done
-            find ./ -type f -name "*.npy" -delete
-        done
-    done
-done

examples/distributed/kunlun/run_pytorch.py

@@ -8,7 +8,6 @@ import time
 import numpy as np
 import onnx
 import os
-import sys
 from onnx.external_data_helper import convert_model_to_external_data
 from onnxsim import simplify
 
@@ -41,10 +40,6 @@ def parse_args():
         default="./",
         help="path to save pytorch model output data"
     )
-    parser.add_argument(
-        "--export_only",
-        action="store_true"
-    )
     args = parser.parse_args()
     print("arg setting: ", args)
     return (
@@ -53,8 +48,7 @@ def parse_args():
         args.length,
         args.export_onnx,
         args.input_dir,
-        args.result_dir,
-        args.export_only
+        args.result_dir
     )
 
 
@@ -83,7 +77,7 @@ def run_pytorch(torch_model, voc_size, batchsize, len, model_name):
     inputs = torch.from_numpy(data).to("cuda")
     torch_model = torch_model.to("cuda")
 
-    n_iter = 10
+    n_iter = 20
     with torch.no_grad():
         for _ in range(10):
             outputs = torch_model(inputs)
@@ -93,11 +87,11 @@ def run_pytorch(torch_model, voc_size, batchsize, len, model_name):
         for _ in range(n_iter):
             torch.cuda.synchronize()
             outputs = torch_model(inputs)
-            #
+            # 
             torch.cuda.synchronize()
     torch.cuda.synchronize()
     end = time.time()
-
+   
     avg_time = (end - begin) / n_iter
     outputs = outputs.last_hidden_state.to("cpu")
     print("outputs abs mean:", abs(np.array(outputs)).mean())
@@ -109,61 +103,33 @@ def run_pytorch(torch_model, voc_size, batchsize, len, model_name):
 
 
 def export_onnx(model_name, model, data, path, extern=False):
-    # torch.onnx.export(model, data, path, verbose=False, do_constant_folding=True)
-
-    if model_name != "llama":
-        onnx_model = onnx.load(path)
+    torch.onnx.export(model, data, path, verbose=False, do_constant_folding=True)
+    onnx_model = onnx.load(path)
+    # onnx_model, check = simplify(onnx_model,
+    #                              skip_shape_inference=True,
+    #                              skipped_optimizers=['eliminate_duplicate_initializer'])
+    if model_name == "gpt2":
+        onnx_model, check = simplify(onnx_model, 
+                                     skip_shape_inference=True,
+                                     skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
+    else :
         onnx_model, check = simplify(onnx_model,
-                                 skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
-                                 # skipped_optimizers=['fuse_qkv'])
-        assert check
-        add_value_info_for_constants(onnx_model)
-        onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
-        if extern:
-            extern_path = path.replace('.onnx', '.pb')
-            if os.path.exists(extern_path):
-                os.remove(extern_path)
-            convert_model_to_external_data(
-                onnx_model,
-                all_tensors_to_one_file=True,
-                location=extern_path.split("/")[-1],
-                size_threshold=1024,
-                convert_attribute=False,
-            )
-        onnx.save(onnx_model, path)
-    else:
-        sys.path.append("onnxsim_large_model")
-        from onnx_utils import set_onnx_input_shape
-        from compress_model import SIZE_1MB, compress_onnx_model, uncompress_onnx_model
-
-        in_model_path = path
-        out_model_path = in_model_path[:-5] + ".sim.onnx"
-
-        onnx_model = onnx.load(in_model_path)
-        print(f"load model from {in_model_path} success")
-
-        size_th_bytes = 1024 * 1024
-        onnx_model, removed_inits = compress_onnx_model(onnx_model, size_th_bytes=size_th_bytes)
-        print("compress model success")
-
-        onnx_model = set_onnx_input_shape(onnx_model, "")
-        tensor_size_threshold = f"1024KB"
-        skipped_optimizers = []
-        skipped_optimizers.append("eliminate_duplicate_initializer")
-        onnx_model, check = simplify(onnx_model, skipped_optimizers=skipped_optimizers,
-                                    tensor_size_threshold=tensor_size_threshold)
-        if not check:
-            raise ValueError(f"simplify compressed model {in_model_path} failed")
-
-        print(f"simplify model success")
-
-        onnx_model = uncompress_onnx_model(onnx_model, removed_inits)
-        print(f"uncompress model success")
-
-        add_value_info_for_constants(onnx_model)
-
-        onnx.save(onnx_model, out_model_path, save_as_external_data=True)
-
+                                     skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
+    assert check
+    add_value_info_for_constants(onnx_model)
+    onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
+    if extern:
+        extern_path = path.replace('.onnx', '.pb')
+        if os.path.exists(extern_path):
+            os.remove(extern_path)
+        convert_model_to_external_data(
+            onnx_model,
+            all_tensors_to_one_file=True,
+            location=extern_path.split("/")[-1],
+            size_threshold=1024,
+            convert_attribute=False,
+        )
+    onnx.save(onnx_model, path)
 
 def add_value_info_for_constants(model : onnx.ModelProto):
     """
@@ -226,20 +192,17 @@ def main():
     global input_dir, result_dir
 
     modelname, batchsize, seqlen, \
-        export_path, input_dir, result_dir, export_only = parse_args()
-
+        export_path, input_dir, result_dir = parse_args()
+    
     model, voc_size = get_model(modelname) # pytorch model
 
     if export_path is not None:
-        os.makedirs(export_path, exist_ok=True)
         filename = "{}_{}_{}.onnx".format(modelname, batchsize, seqlen)
         path = os.path.join(export_path, filename)
         param = torch.zeros((batchsize, seqlen), dtype=torch.int)
         export_onnx(modelname, model, param, path, True) # export pytorch model to onnx model
-        if export_only:
-            return
 
     run_pytorch(model, voc_size, batchsize, seqlen, modelname)
 
 if __name__ == "__main__":
-    main()
+    main()

examples/distributed/kunlun/run_pytorch.sh

@@ -0,0 +1,17 @@
+export HF_ENDPOINT=https://hf-mirror.com
+
+models=("bert" "gpt2")
+batch_size=(1 32)
+seq_len=(100 500)
+nproc=(1 2 4)
+
+for model in "${models[@]}"; do
+    for bs in "${batch_size[@]}"; do
+        for len in "${seq_len[@]}"; do
+            python -m xacc run_pytorch.py --model "$model" --batch_size "$bs" --length "$len" --export_onnx ../models/"$model" > results/"$model"_"$bs"_"$len" 
+                for n in "${nproc[@]}"; do
+                    python kunlun_launch.py --name "$model" --model ../models/"$model"/"$model"_"$bs"_"$len".onnx --nproc_per_node=$n >> results/"$model"_"$bs"_"$len" 
+                done
+        done
+    done
+done

examples/distributed/onnxsim_large_model

@@ -1 +0,0 @@
-Subproject commit cbcf3fbf985a00494b0f136c92eaccd42031bf65

examples/distributed/parallel_opt.py

@@ -80,6 +80,7 @@ def parallel_model(model: ModelProto, tp_world_size: int = 1, tp_rank: int = 0):
 
     def shard_reshape(node: NodeProto):
         # print("reshape", node.name, node.input[0], place[node.input[0]])
+        # import pdb; pdb.set_trace()
         if not is_sharded(node.input[0]):
             return
         in_plc = place[node.input[0]]
@@ -110,6 +111,7 @@ def parallel_model(model: ModelProto, tp_world_size: int = 1, tp_rank: int = 0):
                 s_dim = 0
             elif in_plc.dim == 2:
                 s_dim = 1
+        # import pdb; pdb.set_trace()
         assert s_dim != -1
         assert out_dims[s_dim] % tp_world_size == 0, out_dims
         out_dims[s_dim] //= tp_world_size
@@ -245,3 +247,7 @@ def parallel_model(model: ModelProto, tp_world_size: int = 1, tp_rank: int = 0):
     model = helper.make_model(graph)
     #model = onnx.shape_inference.infer_shapes(model)
     return model
+
+if __name__ == "__main__":
+    model = onnx.load("./models/gpt2/gpt2_1_100.onnx")
+    models = parallel_model(model, 2, 0)

examples/distributed/run_pytorch.py

@@ -10,6 +10,8 @@ import os
 from onnx.external_data_helper import convert_model_to_external_data
 from onnxsim import simplify
 
+torch.backends.cuda.matmul.allow_tf32 = False
+torch.backends.cudnn.allow_tf32 = False
 def parse_args():
     parser = argparse.ArgumentParser(description="Run pytorch gpt2/bert/opt and optionally export onnx.")
     parser.add_argument(
@@ -25,17 +27,14 @@ def parse_args():
         const="./",
         help="whether and where to export onnx file",
     )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], default="fp32", help="data type"
-    )
+    args = parser.parse_args()
     args = parser.parse_args()
     print("arg setting: ", args)
     return (
         args.model,
         args.batch_size,
         args.length,
-        args.export_onnx,
-        args.type,
+        args.export_onnx
     )
 
 
@@ -82,7 +81,7 @@ def run_pytorch(torch_model, voc_size, batchsize, len):
     print("outputs abs mean:", abs(np.array(outputs)).mean())
     print(f"average time: {avg_time}")
     torch.cuda.memory.empty_cache()
-    np.save("test_results", np.array(outputs, dtype=np.float32))
+    np.save("test_results", np.array(outputs))
     print("Save input & output as test_inputs.npy and test_results.npy")
 
 
@@ -165,14 +164,7 @@ def add_value_info_for_constants(model : onnx.ModelProto):
 
 
 def main():
-    torch.backends.cuda.matmul.allow_tf32 = False
-    torch.backends.cudnn.allow_tf32 = False
-    modelname, batchsize, seqlen, export_path, data_type = parse_args()
-    if data_type == "tf32":
-        torch.backends.cuda.matmul.allow_tf32 = True
-    else:
-        os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
-
+    modelname, batchsize, seqlen, export_path = parse_args()
     model, voc_size = get_model(modelname)
     if export_path is not None:
         filename = "{}_{}_{}.onnx".format(modelname, batchsize, seqlen)
@@ -180,8 +172,6 @@ def main():
         param = torch.zeros((batchsize, seqlen), dtype=torch.int)
         export_onnx(model, param, path, True)
 
-    if data_type == "fp16":
-        model = model.half()
     run_pytorch(model, voc_size, batchsize, seqlen)
 
 if __name__ == "__main__":

include/kunlun/kunlun_runtime.h

@@ -21,10 +21,10 @@ class KUNLUNRuntimeObj : public RuntimeObj {
         ctx = xdnn::create_context();
         // 10GB for Longformer
         // size_t longformerNum = 3lu * (1 << 30);
-        size_t workspaceSize = 2llu << 30; // 2 GB
-        KUNLUNPtr wkspacePtr = alloc(workspaceSize);
-        workspace =
-            make_ref<WorkspaceObj<KUNLUNPtr>>(wkspacePtr, workspaceSize);
+        size_t workspaceSize = 3llu << 30; // 3 GB
+                                           // KUNLUNPtr wkspacePtr = alloc(workspaceSize);
+                                           // workspace =
+                                           //     make_ref<WorkspaceObj<KUNLUNPtr>>(wkspacePtr, workspaceSize);
     }
     virtual ~KUNLUNRuntimeObj() {
         KUNLUNPtr wkspacePtr = workspace->getWorkspace();
@@ -33,6 +33,10 @@ class KUNLUNRuntimeObj : public RuntimeObj {
     }
     string toString() const override;
 
+    void init_workspace(KUNLUNPtr workspacePtr) {
+      workspace = make_ref<WorkspaceObj<KUNLUNPtr>>(wkspacePtr, workspaceSize);
+    }
+
     void run(const Graph &graph, bool tune = false,
              bool profiling = false) const;
     // double runEvaluation(const Graph &graph, int nWarmups,
@@ -42,7 +46,7 @@ class KUNLUNRuntimeObj : public RuntimeObj {
     KUNLUNPtr alloc(size_t size) override {
         void *ptr;
         checkKUNLUNError(
-            xpu_malloc((void **)&ptr, size, XPUMemoryKind::XPU_MEM_HBM));
+            xpu_malloc_ex((void **)&ptr, size, XPUMemoryKind::XPU_MEM_MAIN));
         return ptr;
     }
     void dealloc(void *ptr) override { xpu_free(ptr); }

include/kunlun/xccl_communicator.h

@@ -34,8 +34,8 @@ class XcclCommunicatorObj final : public CommunicatorObj {
             auto begin = std::chrono::steady_clock::now();
             while (!std::filesystem::exists(filePath)) {
                 auto now = std::chrono::steady_clock::now();
-                _IT_ASSERT_2(now < begin + std::chrono::seconds(100),
-                             "time limit (100s) exceeded.");
+                _IT_ASSERT_2(now < begin + std::chrono::seconds(10),
+                             "time limit (10s) exceeded.");
                 std::this_thread::sleep_for(std::chrono::milliseconds(100));
             }
             std::ifstream ifs(filePath, std::ios::binary);

src/core/graph.cc

@@ -265,6 +265,11 @@ void GraphObj::dataMalloc(bool useNaiveAllocator, size_t memPoolSize) {
                                      tensorToOffset[tensor.get()]));
         }
     }
+
+    if (runtime->isKUNLUN()) {
+      allocator.heapAlloc(runtime->workspaceSize);
+      runtime->init_workspace(allocator.getHeapPtr());
+    }
 }
 
 Tensor GraphObj::cloneKV(Tensor &tensor) {

src/kernels/bang/cast.cc

@@ -199,24 +199,6 @@ class CastCnnl : public BangKernelWithoutConfig {
                                                    dim.data()));
             NlCastType = CNNL_CAST_UINT32_TO_INT64;
             break;
-        case CastType::Float162Float:
-            checkCnnlError(cnnlSetTensorDescriptor(aDesc, CNNL_LAYOUT_NCHW,
-                                                   CNNL_DTYPE_HALF, dim.size(),
-                                                   dim.data()));
-            checkCnnlError(cnnlSetTensorDescriptor(cDesc, CNNL_LAYOUT_NCHW,
-                                                   CNNL_DTYPE_FLOAT, dim.size(),
-                                                   dim.data()));
-            NlCastType = CNNL_CAST_HALF_TO_FLOAT;
-            break;
-        case CastType::Float2Float16:
-            checkCnnlError(cnnlSetTensorDescriptor(aDesc, CNNL_LAYOUT_NCHW,
-                                                   CNNL_DTYPE_FLOAT, dim.size(),
-                                                   dim.data()));
-            checkCnnlError(cnnlSetTensorDescriptor(cDesc, CNNL_LAYOUT_NCHW,
-                                                   CNNL_DTYPE_HALF, dim.size(),
-                                                   dim.data()));
-            NlCastType = CNNL_CAST_FLOAT_TO_HALF;
-            break;
         default:
             IT_TODO_HALT();
         }

src/kernels/bang/layer_norm.cc

@@ -19,16 +19,14 @@ class LayerNormCnnl : public BangKernelWithoutConfig {
         void *const outputData = (op->getOutput()->getRawDataPtr<void *>());
 
         auto inDims = op->getInputs(0)->getDims();
-        auto fiterDims = op->getInputs(1)->getDims();
         auto outDims = op->getOutput()->getDims();
+        auto fiterDims = op->getOutput(1)->getDims();
 
         float eps = op->getEps();
         const int axis = op->getAxis();
 
-        Shape outMeanDims(outDims);
-        outMeanDims.erase(outMeanDims.begin() + axis);
+        cnnlTensorDescriptor_t inDesc, fiterDesc, outDesc;
 
-        cnnlTensorDescriptor_t inDesc, fiterDesc, outDesc, outMeanDesc;
         checkCnnlError(cnnlCreateTensorDescriptor(&inDesc));
         checkCnnlError(cnnlSetTensorDescriptor(
             inDesc, CNNL_LAYOUT_ARRAY, cnnlDataTypeConvert(op->getDType()),
@@ -41,23 +39,15 @@ class LayerNormCnnl : public BangKernelWithoutConfig {
         checkCnnlError(cnnlSetTensorDescriptor(
             outDesc, CNNL_LAYOUT_ARRAY, cnnlDataTypeConvert(op->getDType()),
             outDims.size(), outDims.data()));
-        checkCnnlError(cnnlCreateTensorDescriptor(&outMeanDesc));
-        checkCnnlError(cnnlSetTensorDescriptor(
-            outMeanDesc, CNNL_LAYOUT_ARRAY, cnnlDataTypeConvert(op->getDType()),
-            outMeanDims.size(), outMeanDims.data()));
         size_t wsSize;
         cnnlGetLayerNormOpWorkspaceSize(context->cnnlHandle(), axis, inDesc,
                                         &wsSize);
         BangPtr wsData = context->getWorkspace(wsSize);
-        size_t meanSize =
-            cnnlGetTensorElementNum(outMeanDesc) * op->getDType().getSize();
-        BangPtr meanData = context->getWorkspace(meanSize);
-        BangPtr rstdData = context->getWorkspace(meanSize);
 
         cnnlStatus_t stat = cnnlLayerNormForward(
             context->cnnlHandle(), inDesc, inputData, axis, fiterDesc,
             scaleData, biasData, eps, wsData, wsSize, outDesc, outputData,
-            outMeanDesc, meanData, rstdData);
+            inDesc, NULL, NULL);
 
         if (stat != CNNL_STATUS_SUCCESS)
             return;

src/kernels/bang/matmul.cc

@@ -66,13 +66,6 @@ class MatmulCnnl : public BangKernelWithoutConfig {
         cnnlSetMatMulDescAttr(bmm_desc, CNNL_MATMUL_DESC_TRANSB, &transB,
                               sizeof(int32_t));
 
-        std::string computeTypeStr = op->getComputeType();
-        if (computeTypeStr == "tf32") {
-            int32_t tf32 = 1;
-            cnnlSetMatMulDescAttr(bmm_desc, CNNL_MATMUL_ALLOW_TF32, &tf32,
-                                  sizeof(int32_t));
-        }
-
         cnnlMatMulAlgo_t bmm_algo;
         cnnlMatMulAlgoCreate(&bmm_algo);
 

src/kernels/kunlun/all_reduce.cc

@@ -20,9 +20,15 @@ class AllReduceXCCL : public KUNLUNKernelWithoutConfig {
         BKCLContext_t comm =
             dynamic_cast<XcclCommunicatorObj &>(context->getCommunicator())
                 .getXcclComm();
+        // double t = timeit(
+        // [&]() {
         checkXcclError(bkcl_all_reduce(comm, input, output, count,
                                        BKCLDataType::BKCL_FLOAT, getRedOp(),
                                        0));
+        // },
+        // [&]() { context->sync(); });
+        // std::cout << "Time consuming for " << op->getInputs(0)->size() << "
+        // size is " << t << std::endl;
     }
     virtual BKCLOp getRedOp() const = 0;
 };

src/kernels/kunlun/element_wise.cc

@@ -97,14 +97,11 @@ class DivXdnn : public KUNLUNKernelWithoutConfig {
         auto aDim = op->getInputs(0)->getDims();
         auto bSize = op->getInputs(1)->size();
         auto bDim = op->getInputs(1)->getDims();
+        auto dtype = op->getDType();
 
-        // op input a, b is scalar while aDim and b Dim is empty
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        if (aDim.size() == 0) {
-            aDim.push_back(1);
-        }
 
         if (aSize == bSize) {
             // Do ElementWise Sub with no broadcast
@@ -112,9 +109,23 @@ class DivXdnn : public KUNLUNKernelWithoutConfig {
                                               (float *)aData, (float *)bData,
                                               (float *)cData, aSize));
         } else {
-            checkKUNLUNError(xdnn::broadcast_div<float>(
-                context->KUNLUNHandle(), (float *)aData, (float *)bData,
-                (float *)cData, aDim, bDim));
+            // Do broadcast div
+            Shape aligned = infer_broadcast(aDim, bDim);
+            if (aligned == aDim) {
+                // BData need to be broadcasted
+                checkKUNLUNError(xdnn::broadcast_div<float>(
+                    context->KUNLUNHandle(), (float *)aData, (float *)bData,
+                    (float *)cData, aDim, bDim));
+            } else {
+                // Use workspace to broadcast aData
+                KUNLUNPtr wks = context->getWorkspace(bSize * dtype.getSize());
+                checkKUNLUNError(xdnn::broadcast<float>(
+                    context->KUNLUNHandle(), (float *)aData, (float *)wks, aDim,
+                    bDim));
+                checkKUNLUNError(xdnn::div<float>(context->KUNLUNHandle(),
+                                                  (float *)wks, (float *)bData,
+                                                  (float *)cData, bSize));
+            }
         }
         return;
     }

src/kernels/kunlun/unary.cc

@@ -570,7 +570,6 @@ REGISTER_KERNEL(Device::KUNLUN, OpType::Reciprocal, ReciprocalXdnn,
 REGISTER_KERNEL(Device::KUNLUN, OpType::Reshape, CopyXdnn, "Reshape_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Flatten, CopyXdnn, "Flatten_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Identity, CopyXdnn, "Identity_xdnn");
-REGISTER_KERNEL(Device::KUNLUN, OpType::Squeeze, CopyXdnn, "Squeeze_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Abs, AbsXdnn, "Abs_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Atan, ATanXdnn, "Atan_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Log, LogXdnn, "Log_xdnn");