Support deepspeed AutoTP (#9230)

* Support deepspeed

* add test script

* refactor convert

* refine example

* refine

* refine example

* fix style

* refine example and adapte latest ipex

* fix style

python/llm/example/GPU/Deepspeed-AutoTP/deepspeed_autotp.py

@@ -0,0 +1,103 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import os
+import torch
+import transformers
+import deepspeed
+
+local_rank = int(os.getenv("LOCAL_RANK", "0"))
+world_size = int(os.getenv("WORLD_SIZE", "1"))
+
+from bigdl.llm import optimize_model
+
+import torch
+import intel_extension_for_pytorch as ipex
+import time
+import argparse
+
+from transformers import AutoModelForCausalLM  # export AutoModelForCausalLM from transformers so that deepspeed use it
+from transformers import LlamaTokenizer, AutoTokenizer
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="meta-llama/Llama-2-7b-chat-hf",
+                        help='The huggingface repo id for the Llama2 (e.g. `meta-llama/Llama-2-7b-chat-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--prompt', type=str, default="Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun",
+                        help='Prompt to infer')
+    parser.add_argument('--n-predict', type=int, default=32,
+                        help='Max tokens to predict')
+
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_path
+
+    model = AutoModelForCausalLM.from_pretrained(args.repo_id_or_model_path,
+                                                 low_cpu_mem_usage=True,
+                                                 torch_dtype=torch.float16,
+                                                 trust_remote_code=True,
+                                                 use_cache=True)
+
+    model = deepspeed.init_inference(
+        model,
+        mp_size=world_size,
+        dtype=torch.float16,
+        replace_method="auto",
+    )
+
+    # move model to cpu and use bigdl-llm `optimize_model` to convert the
+    # model into optimized low bit format
+    # convert the rest of the model into float16 to reduce allreduce traffic
+    model = optimize_model(model.module.to(f'cpu'), low_bit='sym_int4').to(torch.float16)
+
+    # move model back to xpu
+    model = model.to(f'xpu:{local_rank}')
+
+    print(model)
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+    # Generate predicted tokens
+    with torch.inference_mode():
+        # prompt = get_prompt(args.prompt, [], system_prompt=DEFAULT_SYSTEM_PROMPT)
+        prompt = args.prompt
+        # input_str = f"Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{prompt}\n\n### Response:\n"
+        input_ids = tokenizer.encode(prompt, return_tensors="pt").to(f'xpu:{local_rank}')
+        # ipex model needs a warmup, then inference time can be accurate
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict,
+                                use_cache=True)
+
+        # start inference
+        st = time.time()
+        # if your selected model is capable of utilizing previous key/value attentions
+        # to enhance decoding speed, but has `"use_cache": false` in its model config,
+        # it is important to set `use_cache=True` explicitly in the `generate` function
+        # to obtain optimal performance with BigDL-LLM INT4 optimizations
+        output = model.generate(input_ids,
+                                do_sample=False,
+                                max_new_tokens=args.n_predict)
+        torch.xpu.synchronize()
+        end = time.time()
+        if local_rank == 0:
+            output = output.cpu()
+            output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+            print(f'Inference time: {end-st} s')
+            print('-'*20, 'Prompt', '-'*20)
+            print(prompt)
+            print('-'*20, 'Output', '-'*20)
+            print(output_str)

python/llm/example/GPU/Deepspeed-AutoTP/run.sh

@@ -0,0 +1,12 @@
+source bigdl-llm-init -t -g
+export MASTER_ADDR=127.0.0.1
+export CCL_ZE_IPC_EXCHANGE=sockets
+if [[ -n $OMP_NUM_THREADS ]]; then
+    export OMP_NUM_THREADS=$(($OMP_NUM_THREADS / 4))
+else
+    export OMP_NUM_THREADS=$(($(nproc) / 4))
+fi
+torchrun --standalone \
+         --nnodes=1 \
+         --nproc-per-node 4 \
+         deepspeed_autotp.py --repo-id-or-model-path "meta-llama/Llama-2-7b-hf"

python/llm/src/bigdl/llm/transformers/convert.py

@@ -45,6 +45,42 @@ from bigdl.llm.ggml.quantize import ggml_tensor_qtype
 from .utils import logger
 
 
+def is_deepspeed_available():
+    return importlib.util.find_spec("deepspeed") is not None
+
+
+def is_linear_module(module):
+
+    in_features = None
+    out_features = None
+    mp_group = None
+
+    if isinstance(module, nn.Linear):
+        in_features = module.in_features
+        out_features = module.out_features
+        mp_group = None
+        result = True
+    else:
+        if is_deepspeed_available():
+            from deepspeed.module_inject.layers import LinearLayer, LinearAllreduce
+            if isinstance(module, LinearLayer):
+                in_features = module.weight.shape[1]
+                out_features = module.weight.shape[0]
+                mp_group = None
+                result = True
+            elif isinstance(module, LinearAllreduce):
+                in_features = module.weight.shape[1]
+                out_features = module.weight.shape[0]
+                mp_group = module.mp_group
+                result = True
+            else:
+                result = False
+        else:
+            result = False
+
+    return result, (in_features, out_features, mp_group)
+
+
 def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
                                  current_key_name=None, convert_shape_only=False):
     from bigdl.llm.transformers.low_bit_linear import LowBitLinear, FP4Params, FP16Linear
@@ -54,17 +90,20 @@ def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
         if current_key_name is None:
             current_key_name = []
 
-        if isinstance(module, nn.Linear) and name not in modules_to_not_convert:
+        is_linear, linear_args = is_linear_module(module)
+        if is_linear and name not in modules_to_not_convert:
             # Check if the current key is not in the `modules_to_not_convert`
             if not any(key in ".".join(current_key_name) for key in modules_to_not_convert):
+                in_features, out_features, mp_group = linear_args
                 with init_empty_weights():
                     new_linear = None
                     if qtype != ggml_tensor_qtype["fp16"]:
                         new_linear = LowBitLinear(
-                            module.in_features,
+                            in_features,
-                            module.out_features,
+                            out_features,
                             qtype,
                             module.bias is not None,
+                            mp_group=mp_group,
                         )
 
                         device_type = module.weight.data.device.type
@@ -82,10 +121,11 @@ def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
                         if module.in_features in [4096, 11008]:
                             # esimd fp16 path
                             new_linear = FP16Linear(
-                                module.in_features,
+                                in_features,
-                                module.out_features,
+                                out_features,
                                 qtype,
                                 module.bias is not None,
+                                mp_group=mp_group,
                             )
                             device_type = module.weight.data.device.type
 

python/llm/src/bigdl/llm/transformers/low_bit_linear.py

@@ -328,7 +328,7 @@ class MatMulLowBit(torch.autograd.Function):
 
 class LowBitLinear(nn.Linear):
     def __init__(self, input_features, output_features, qtype, bias=True,
-                 conver_to_half=True):
+                 conver_to_half=True, mp_group=None):
         super().__init__(input_features, output_features, bias)
         self.weight = FP4Params(self.weight.data,
                                 requires_grad=False,
@@ -339,6 +339,7 @@ class LowBitLinear(nn.Linear):
         self.weight_length = self.out_len * self.in_len
         self.qtype = qtype
         self.conver_to_half = conver_to_half
+        self.mp_group = mp_group
 
     def forward(self, x: torch.Tensor):
         if self.bias is not None and self.bias.dtype != x.dtype:
@@ -378,6 +379,9 @@ class LowBitLinear(nn.Linear):
                                                      input_seq_size)
             new_shape = x_shape[:-1] + (self.out_len,)
             result = result.view(new_shape)
+            if self.mp_group is not None:
+                from deepspeed import comm as dist
+                dist.inference_all_reduce(result, group=self.mp_group)
             if self.bias is not None:
                 result += self.bias
         else:
@@ -400,7 +404,7 @@ class LowBitLinear(nn.Linear):
 
 class FP16Linear(nn.Linear):
     def __init__(self, input_features, output_features, qtype, bias=True,
-                 conver_to_half=True):
+                 conver_to_half=True, mp_group=None):
         super().__init__(input_features, output_features, bias)
         self.in_len = input_features
         self.out_len = output_features
@@ -408,6 +412,7 @@ class FP16Linear(nn.Linear):
         self.weight_length = self.out_len * self.in_len
         self.qtype = qtype
         self.conver_to_half = conver_to_half
+        self.mp_group = mp_group
 
     def forward(self, x: torch.Tensor):
         if self.bias is not None and self.bias.dtype != x.dtype:
@@ -442,6 +447,9 @@ class FP16Linear(nn.Linear):
 
         new_shape = x_shape[:-1] + (self.out_len,)
         result = result.view(new_shape)
+        if self.mp_group is not None:
+            from deepspeed import comm as dist
+            dist.inference_all_reduce(result, group=self.mp_group)
         if self.bias is not None:
             result += self.bias
 

python/llm/src/bigdl/llm/transformers/models/llama.py

@@ -32,6 +32,7 @@
 # limitations under the License.
 
 import torch
+import importlib
 import torch.nn as nn
 from typing import Optional, Tuple
 import math
@@ -58,10 +59,27 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
 
 
+def get_ipex_version():
+
+    if importlib.util.find_spec("intel_extension_for_pytorch") is not None:
+        import intel_extension_for_pytorch as ipex
+        return ipex.__version__
+    else:
+        return None
+
+
+ipex_version = get_ipex_version()
+
+
 def llama_rms_norm_forward(self, hidden_states):
     if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
+        if ipex_version == "2.0.110+xpu":
             hidden_states, _ = torch.ops.torch_ipex.rms_norm(hidden_states,
                                                              [self.weight.size(0)], self.weight)
+        else:
+            hidden_states, _ = torch.ops.torch_ipex.rms_norm(hidden_states,
+                                                             [self.weight.size(0)], self.weight,
+                                                             self.variance_epsilon)
     else:
         input_dtype = hidden_states.dtype
         hidden_states = hidden_states.to(torch.float32)