[Hardware][TPU] Initial support of model parallelism with single worker using SPMD (vllm-project#18011)

Signed-off-by: Siyuan Liu <lsiyuan@google.com>
Co-authored-by: Hossein Sarshar <hossein.sarshar@gmail.com>
Co-authored-by: Chengji Yao <chengjiyao@google.com>
Signed-off-by: minpeter <kali2005611@gmail.com>

Author: 3 people

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -155,6 +155,10 @@ run_and_track_test 12 "test_moe_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
 run_and_track_test 13 "test_lora.py" \
     "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
+run_and_track_test 14 "test_tpu_qkv_linear.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
+run_and_track_test 15 "test_spmd_model_weight_loading.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
 
 # After all tests have been attempted, exit with the overall status.
 if [ "$overall_script_exit_code" -ne 0 ]; then

examples/offline_inference/tpu.py

@@ -1,5 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import argparse
+import os
+
 from vllm import LLM, SamplingParams
 
 prompts = [
@@ -18,14 +21,28 @@
 
 
 def main():
+    parser = argparse.ArgumentParser(description="TPU offline inference example")
+    parser.add_argument("--use-spmd", action="store_true", help="Enable SPMD mode")
+    args = parser.parse_args()
+
+    llm_args = {
+        "model": "Qwen/Qwen2-1.5B-Instruct",
+        "max_num_batched_tokens": 64,
+        "max_num_seqs": 4,
+        "max_model_len": 128,
+    }
+    if args.use_spmd:
+        os.environ["VLLM_XLA_USE_SPMD"] = "1"
+        # Can only hardcode the number of chips for now.
+        # calling xr.global_runtime_device_count() beforeing init SPMD env in
+        # torch_xla will mess up the distributed env.
+        llm_args["tensor_parallel_size"] = 8
+        # Use Llama, for num_kv_heads = 8.
+        llm_args["model"] = "meta-llama/Llama-3.1-8B-Instruct"
+
     # Set `enforce_eager=True` to avoid ahead-of-time compilation.
     # In real workloads, `enforace_eager` should be `False`.
-    llm = LLM(
-        model="Qwen/Qwen2-1.5B-Instruct",
-        max_num_batched_tokens=64,
-        max_num_seqs=4,
-        max_model_len=128,
-    )
+    llm = LLM(**llm_args)
     outputs = llm.generate(prompts, sampling_params)
     print("-" * 50)
     for output, answer in zip(outputs, answers):

tests/v1/tpu/test_spmd_model_weight_loading.py

@@ -0,0 +1,67 @@
+# SPDX-License-Identifier: Apache-2.0
+import gc
+import tempfile
+
+import numpy as np
+import pytest
+import torch_xla.distributed.spmd as xs
+import torch_xla.runtime as xr
+
+from vllm.config import set_current_vllm_config
+from vllm.distributed.parallel_state import (ensure_model_parallel_initialized,
+                                             init_distributed_environment)
+from vllm.engine.arg_utils import EngineArgs
+from vllm.model_executor.model_loader.tpu import TPUModelLoader
+
+
+def _setup_environment(model):
+    engine_args = EngineArgs(model=model, )
+    vllm_config = engine_args.create_engine_config()
+    with set_current_vllm_config(vllm_config):
+        temp_file = tempfile.mkstemp()[1]
+        init_distributed_environment(
+            1,
+            0,
+            local_rank=0,
+            distributed_init_method=f"file://{temp_file}",
+            backend="gloo")
+        # Under single worker mode, full model is init first and then
+        # partitioned using GSPMD.
+        ensure_model_parallel_initialized(1, 1)
+    return vllm_config
+
+
+MESH = None
+
+
+def _get_spmd_mesh():
+    global MESH
+    if MESH is None:
+        xr.use_spmd()
+        num_devices = xr.global_runtime_device_count()
+        mesh_shape = (num_devices, 1)
+        device_ids = np.array(range(num_devices))
+        MESH = xs.Mesh(device_ids, mesh_shape, ('x', 'y'))
+    return MESH
+
+
+@pytest.mark.parametrize("model", [
+    "Qwen/Qwen2-1.5B-Instruct",
+    "meta-llama/Llama-3.1-8B-Instruct",
+    "meta-llama/Llama-3.1-70B-Instruct",
+])
+def test_tpu_model_loader(model):
+    # Skip the 70B test if there are less than 8 chips
+    # TODO: Query using torch xla API, the query API is not working
+    # with SPMD now. However, This test is running under SPMD mode.
+    if '70B' in model and xr.global_runtime_device_count() < 8:
+        pytest.skip(
+            "Skipping 70B model if the TPU VM has less than 8 chips to \
+                     avoid OOM.")
+
+    vllm_config = _setup_environment(model)
+    loader = TPUModelLoader(load_config=vllm_config.load_config)
+    mesh = _get_spmd_mesh()
+    model = loader.load_model(vllm_config, vllm_config.model_config, mesh)
+    del model
+    gc.collect()

tests/v1/tpu/test_tpu_qkv_linear.py

@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: Apache-2.0
+import tempfile
+
+import numpy as np
+import pytest
+import torch
+import torch_xla.distributed.spmd as xs
+import torch_xla.runtime as xr
+
+from vllm.config import set_current_vllm_config
+from vllm.distributed.parallel_state import (ensure_model_parallel_initialized,
+                                             init_distributed_environment)
+from vllm.distributed.tpu_distributed_utils import XlaQKVParallelLinear
+from vllm.engine.arg_utils import EngineArgs
+from vllm.model_executor.layers.linear import QKVParallelLinear
+
+
+@pytest.fixture(autouse=True)
+def setup_environment():
+    # This is a fake config used for init dist env.
+    # QKVParallelLinear needs dist env to be initialized.
+    engine_args = EngineArgs(
+        model="Qwen/Qwen2-1.5B-Instruct",
+        max_model_len=64,
+        max_num_batched_tokens=64,
+        max_num_seqs=4,
+    )
+
+    vllm_config = engine_args.create_engine_config()
+
+    with set_current_vllm_config(vllm_config):
+        temp_file = tempfile.mkstemp()[1]
+        init_distributed_environment(
+            1,
+            0,
+            local_rank=0,
+            distributed_init_method=f"file://{temp_file}",
+            backend="gloo")
+        ensure_model_parallel_initialized(1, 1)
+        yield
+
+
+MESH = None
+
+
+def _get_spmd_mesh():
+    global MESH
+    if MESH is None:
+        xr.use_spmd()
+        num_devices = xr.global_runtime_device_count()
+        mesh_shape = (num_devices, 1)
+        device_ids = np.array(range(num_devices))
+        MESH = xs.Mesh(device_ids, mesh_shape, ('x', 'y'))
+    return MESH
+
+
+@pytest.mark.parametrize("bias", [False, True])
+# `xr.use_spmd()` will set a global state, and this state is not reversible.
+# Therefore, non-SPMD tests should be run before SPMD tests.
+@pytest.mark.parametrize("mesh", [None, _get_spmd_mesh()])
+@pytest.mark.parametrize("device", ['cpu', 'xla'])
+@torch.no_grad()
+def test_xla_qkv_linear(bias, mesh, device):
+    torch.manual_seed(123)
+
+    qkv_linear = QKVParallelLinear(
+        hidden_size=4096,
+        head_size=128,
+        total_num_heads=32,
+        total_num_kv_heads=8,
+        bias=bias,
+        params_dtype=torch.bfloat16,
+        return_bias=False,
+    )
+
+    qkv_linear.weight.data = torch.rand_like(qkv_linear.weight.data) / 10
+    if bias:
+        qkv_linear.bias.data = torch.rand_like(qkv_linear.bias.data)
+
+    xla_qkv_linear = XlaQKVParallelLinear(qkv_linear, mesh=mesh)
+
+    qkv_linear = qkv_linear.to(device)
+    xla_qkv_linear = xla_qkv_linear.to(device)
+    input_tensor = torch.rand(10, 4096, dtype=torch.bfloat16) / 10
+    input_tensor = input_tensor.to(device)
+
+    output = qkv_linear(input_tensor)
+    xla_output = xla_qkv_linear(input_tensor)
+    assert torch.allclose(output.cpu(), xla_output.cpu())

vllm/config.py

@@ -1901,6 +1901,8 @@ def __post_init__(self) -> None:
             if current_platform.is_neuron():
                 # neuron uses single process to control multiple devices
                 backend = "uni"
+            elif current_platform.is_tpu() and envs.VLLM_XLA_USE_SPMD:
+                backend = "uni"
             elif (current_platform.is_cuda()
                   and cuda_device_count_stateless() < self.world_size):
                 if not ray_found:

vllm/distributed/tpu_distributed_utils.py

@@ -0,0 +1,177 @@
+# SPDX-License-Identifier: Apache-2.0
+from collections import OrderedDict
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch_xla.distributed.spmd as xs
+from torch.nn.parameter import Parameter
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+
+logger = init_logger(__name__)
+
+
+class XlaQKVParallelLinear(nn.Module):
+
+    def __init__(self,
+                 qkv_linear: nn.Module,
+                 mesh: Optional["xs.Mesh"] = None):
+        super().__init__()
+        assert isinstance(qkv_linear, QKVParallelLinear)
+        self.skip_bias_add = qkv_linear.skip_bias_add
+        self.return_bias = qkv_linear.return_bias
+        assert qkv_linear.tp_size == 1, "TP > 1 is only supported under SPMD."
+
+        self.q_weight: Parameter
+        self.k_weight: Parameter
+        self.v_weight: Parameter
+        self.q_bias: Optional[Parameter]
+        self.k_bias: Optional[Parameter]
+        self.v_bias: Optional[Parameter]
+        self._load_weights_from_qkv_linear(qkv_linear)
+        if mesh is not None:
+            self._shard_weight(mesh)
+
+    def _shard_weight(self, mesh: "xs.Mesh"):
+        self.q_weight = Parameter(self.q_weight.to('xla'), requires_grad=False)
+        self.k_weight = Parameter(self.k_weight.to('xla'), requires_grad=False)
+        self.v_weight = Parameter(self.v_weight.to('xla'), requires_grad=False)
+        xs.mark_sharding(self.q_weight, mesh, ('x', None))
+        xs.mark_sharding(self.k_weight, mesh, ('x', None))
+        xs.mark_sharding(self.v_weight, mesh, ('x', None))
+        if self.q_bias is not None:
+            assert self.k_bias is not None and self.v_bias is not None, \
+                "QKVParallelLinear should have q, k, and v biases together."
+            self.q_bias = Parameter(self.q_bias.to('xla'), requires_grad=False)
+            xs.mark_sharding(self.q_bias, mesh, ('x', ))
+            self.k_bias = Parameter(self.k_bias.to('xla'), requires_grad=False)
+            xs.mark_sharding(self.k_bias, mesh, ('x', ))
+            self.v_bias = Parameter(self.v_bias.to('xla'), requires_grad=False)
+            xs.mark_sharding(self.v_bias, mesh, ('x', ))
+
+    def _load_weights_from_qkv_linear(self, qkv_linear: nn.Module):
+        q_proj_size, k_proj_size, _ = qkv_linear.output_sizes
+        # The weight of qkv linear is a concatenation of q, k, and v weights
+        # along the output dimension.
+        qkv_weight = qkv_linear.weight.data.cpu()
+        q_weight = Parameter(qkv_weight[:q_proj_size], requires_grad=False)
+        k_weight = Parameter(qkv_weight[q_proj_size:q_proj_size + k_proj_size],
+                             requires_grad=False)
+        v_weight = Parameter(qkv_weight[q_proj_size + k_proj_size:],
+                             requires_grad=False)
+        self.register_parameter("q_weight", q_weight)
+        self.register_parameter("k_weight", k_weight)
+        self.register_parameter("v_weight", v_weight)
+
+        if qkv_linear.bias is not None:
+            q_bias = Parameter(qkv_linear.bias[:q_proj_size],
+                               requires_grad=False)
+            k_bias = Parameter(qkv_linear.bias[q_proj_size:q_proj_size +
+                                               k_proj_size],
+                               requires_grad=False)
+            v_bias = Parameter(qkv_linear.bias[q_proj_size + k_proj_size:],
+                               requires_grad=False)
+            self.register_parameter("q_bias", q_bias)
+            self.register_parameter("k_bias", k_bias)
+            self.register_parameter("v_bias", v_bias)
+        else:
+            self.register_parameter("q_bias", None)
+            self.register_parameter("k_bias", None)
+            self.register_parameter("v_bias", None)
+
+    def forward(self, input):
+        # Same forward functionality as QKVParallelLinear, but doing qkv porj
+        # separately.
+        q_bias = self.q_bias if not self.skip_bias_add else None
+        k_bias = self.k_bias if not self.skip_bias_add else None
+        v_bias = self.v_bias if not self.skip_bias_add else None
+        q_proj = F.linear(input, self.q_weight, q_bias)
+        k_proj = F.linear(input, self.k_weight, k_bias)
+        v_proj = F.linear(input, self.v_weight, v_bias)
+        # The q/k/v projections will be split outside of the QKVParallelLinear.
+        # Because we are replacing XlaQKVParallelLinear with the
+        # QKVParallelLinear, we need to concatenate q, k, and v projections to
+        # match the output shape of the QKVParallelLinear implementation even if
+        # it seems to be redundant.
+        # The concat and the following split will be noop, and should be
+        # optimized away by the compiler.
+        qkv_proj = torch.cat([q_proj, k_proj, v_proj], dim=-1)
+        output_bias = torch.cat([q_bias, k_bias, v_bias], dim=-1) if \
+                            self.skip_bias_add else None
+        if not self.return_bias:
+            return qkv_proj
+        return qkv_proj, output_bias
+
+
+def partition_column_parallel_linear(layer: torch.nn.Module,
+                                     mesh: xs.Mesh) -> torch.nn.Module:
+    assert isinstance(layer, ColumnParallelLinear)
+    xs.mark_sharding(layer.weight, mesh, ('x', None))
+    logger.debug("Applied column-parallel sharding to %s", layer)
+    return layer
+
+
+def partition_row_parallel_linear(layer: torch.nn.Module,
+                                  mesh: xs.Mesh) -> torch.nn.Module:
+    assert isinstance(layer, RowParallelLinear)
+    xs.mark_sharding(layer.weight, mesh, (None, 'x'))
+    logger.debug("Applied row-parallel sharding to %s", layer)
+    return layer
+
+
+def partition_qkv_parallel_linear(layer: torch.nn.Module,
+                                  mesh: xs.Mesh) -> torch.nn.Module:
+    assert isinstance(layer, QKVParallelLinear)
+    xla_layer = XlaQKVParallelLinear(layer, mesh)
+    logger.debug("Applied qkv parallel sharding to %s", layer)
+    return xla_layer
+
+
+MODULE_TYPE_TO_WRAPPING_FUNC = OrderedDict([
+    ("QKVParallelLinear", partition_qkv_parallel_linear),
+    ("ColumnParallelLinear", partition_column_parallel_linear),
+    ("RowParallelLinear", partition_row_parallel_linear),
+])
+
+
+def get_fqn(module):
+    # Get the fully qualified name of the module
+    return module.__class__.__qualname__
+
+
+def shard_model(model: torch.nn.Module, mesh: "xs.Mesh") -> None:
+    """
+    Recursively check a PyTorch model and apply appropriate sharding based on 
+    the MODULE_TYPE_TO_WRAPPING_FUNC mapping.
+    
+    Args:
+        model: torch.nn.Module to process
+        mesh: An XLA SPMD mesh object used for sharding
+    """
+
+    def _process_module(module, name=None, parent=None):
+        for module_type, wrapping_func in MODULE_TYPE_TO_WRAPPING_FUNC.items():
+            if get_fqn(module) == module_type:
+                wrapped_module = wrapping_func(module, mesh)
+
+                assert parent is not None and name is not None, (
+                    "Top Level module is not expected to be wrapped.")
+                if wrapped_module is not module:
+                    # Wrapped module and module are different py object.
+                    # The original module should be replaced by the
+                    # wrapped_module.
+                    logger.debug("replace %s with %s", module, wrapped_module)
+                    setattr(parent, name, wrapped_module)
+
+                module = wrapped_module
+                break
+
+        for child_name, child_module in list(module.named_children()):
+            _process_module(child_module, child_name, module)
+
+    _process_module(model)