Add Tensor Parallel to torch_native_llama (#1876)

python/sglang/bench_latency.py

@@ -220,8 +220,7 @@ def prepare_synthetic_inputs_for_latency_test(batch_size, input_len):
     return reqs
 
 
-@torch.inference_mode()
-def extend(reqs, model_runner):
+def _extend(reqs, model_runner):
     batch = ScheduleBatch.init_new(
         reqs=reqs,
         req_to_token_pool=model_runner.req_to_token_pool,
@@ -237,8 +236,15 @@ def extend(reqs, model_runner):
     return next_token_ids, logits_output.next_token_logits, batch
 
 
-@torch.inference_mode()
-def decode(input_token_ids, batch, model_runner):
+def extend(reqs, model_runner):
+    # Disable inference mode for now when torch TP is applied. We can remove
+    # this workaround once DTensor adds support for inference mode.
+    use_inf_mode = not model_runner.torch_tp_applied
+    with torch.inference_mode(use_inf_mode):
+        return _extend(reqs, model_runner)
+
+
+def _decode(input_token_ids, batch, model_runner):
     batch.output_ids = input_token_ids
     batch.prepare_for_decode()
     model_worker_batch = batch.get_model_worker_batch()
@@ -248,6 +254,14 @@ def decode(input_token_ids, batch, model_runner):
     return next_token_ids, logits_output.next_token_logits
 
 
+def decode(input_token_ids, batch, model_runner):
+    # Disable inference mode for now when torch TP is applied. We can remove
+    # this workaround once DTensor adds support for inference mode.
+    use_inf_mode = not model_runner.torch_tp_applied
+    with torch.inference_mode(use_inf_mode):
+        return _decode(input_token_ids, batch, model_runner)
+
+
 def correctness_test(
     server_args,
     port_args,

python/sglang/srt/model_executor/model_runner.py

@@ -148,6 +148,15 @@ class ModelRunner:
         min_per_gpu_memory = self.init_torch_distributed()
         self.sampler = Sampler()
         self.load_model()
+
+        # Apply torch TP if model supports it
+        supports_torch_tp = getattr(self.model, "supports_torch_tp", False)
+        if self.tp_size > 1 and supports_torch_tp:
+            self.apply_torch_tp()
+            self.torch_tp_applied = True
+        else:
+            self.torch_tp_applied = False
+
         if server_args.lora_paths is not None:
             self.init_lora_manager()
         self.init_memory_pool(
@@ -551,6 +560,13 @@ class ModelRunner:
         logger.info("Capture cuda graph begin. This can take up to several minutes.")
         self.cuda_graph_runner = CudaGraphRunner(self)
 
+    def apply_torch_tp(self):
+        logger.info(f"Enabling torch tensor parallelism on {self.tp_size} devices.")
+        from sglang.srt.model_parallel import tensor_parallel
+
+        device_mesh = torch.distributed.init_device_mesh(self.device, (self.tp_size,))
+        tensor_parallel(self.model, device_mesh)
+
     def forward_decode(self, forward_batch: ForwardBatch):
         if self.cuda_graph_runner and self.cuda_graph_runner.can_run(forward_batch):
             return self.cuda_graph_runner.replay(forward_batch)

python/sglang/srt/model_parallel.py

+"""
+Common utilities for torch model parallelism.
+"""
+
+from typing import Optional, Sequence
+
+import torch
+from torch.distributed.device_mesh import DeviceMesh
+
+try:
+    from torch.distributed.tensor import DTensor, Shard
+except ImportError:
+    # torch 2.4 or older
+    from torch.distributed._tensor import DTensor, Shard
+
+from torch.distributed._functional_collectives import AsyncCollectiveTensor
+from torch.distributed.tensor.parallel import (
+    ColwiseParallel,
+    RowwiseParallel,
+    parallelize_module,
+)
+
+
+class ColwiseParallelSharded(ColwiseParallel):
+    """
+    A version of ColwiseParallel where the local weight has been already
+    sharded.  This is used for the fused wqkv case, where during loading, we
+    already sharded wq, wk, wv before fusing them.
+    """
+
+    # Override the _partition_linear_fn in ColwiseParallel
+    def _partition_linear_fn(self, name, module, device_mesh):
+        # colwise shard weight/bias to Shard(0), weight be Shard(0)
+        # means Colwise as Linear is input * weight^T + bias, where
+        # weight would become Shard(1)
+        for name, param in module.named_parameters():
+            dtensor = DTensor.from_local(param, device_mesh, [Shard(0)])
+            dist_param = torch.nn.Parameter(dtensor, requires_grad=False)
+            module.register_parameter(name, dist_param)
+
+
+class RowwiseParallelMaybeWait(RowwiseParallel):
+    """
+    A version of RowwiseParallel that waits for the output (establish dependency
+    between comm stream and compute stream in CUDA sense) before going into the
+    next op. This is needed to workaround the current interaction between
+    AsyncCollectiveTensor and custom ops, such as `class RMSNorm(CustomOp)`.
+    """
+
+    @staticmethod
+    def _prepare_output_fn(output_layouts, use_local_output, mod, outputs, device_mesh):
+        outputs = super(
+            RowwiseParallelMaybeWait, RowwiseParallelMaybeWait
+        )._prepare_output_fn(
+            output_layouts, use_local_output, mod, outputs, device_mesh
+        )
+        # wait for the output to be ready
+        if isinstance(outputs, AsyncCollectiveTensor):
+            return outputs.wait()
+        else:
+            return outputs
+
+
+def tensor_parallel(
+    module: torch.nn.Module,
+    device_mesh: Optional[DeviceMesh] = None,
+):
+    """
+    Tensor parallelize the model across the given device mesh.
+    Args:
+        module (`torch.nn.Module`):
+            The module to tensor parallelize.
+        device_mesh (`torch.distributed.DeviceMesh`):
+            The device mesh to use for tensor parallelism.
+    """
+
+    # Tensor parallelize a nn.Module based on the `_tp_plan` attribute of the module.
+    # No op if `_tp_plan` attribute does not exist under the module.
+    # This is a helper function to be used with `model.apply` to recursively
+    # parallelize a model.
+    def tplize(mod: torch.nn.Module) -> None:
+        tp_plan = getattr(mod, "_tp_plan", None)
+        if tp_plan is None:
+            return
+        for child_name, tp_style in tp_plan.items():
+            submod = mod.get_submodule(child_name)
+            if tp_style == "Colwise":
+                parallelize_module(submod, device_mesh, ColwiseParallel())
+            elif tp_style == "Rowwise":
+                parallelize_module(submod, device_mesh, RowwiseParallelMaybeWait())
+            elif tp_style == "Colwise_Sharded":
+                parallelize_module(submod, device_mesh, ColwiseParallelSharded())
+            else:
+                raise ValueError(f"Unknown TP style {tp_style}")
+
+    # `apply` is a native method of `nn.Module` that recursively applies a
+    # function to every submodule.
+    module.apply(tplize)

python/sglang/srt/models/torch_native_llama.py

@@ -17,6 +17,31 @@ limitations under the License.
 # https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/llama.py#L1
 """Inference-only LLaMA model compatible with HuggingFace weights."""
 
+# PyTorch Tensor Parallel Available for This Model
+"""
+This model supports tensor parallelism (TP) using the PyTorch tensor parallel package.
+Reference: https://pytorch.org/docs/stable/distributed.tensor.parallel.html
+
+Here is a quick example to enable TP:
+```python
+from sglang.srt.model_parallel import tensor_parallel
+
+device_mesh = torch.distributed.init_device_mesh("cuda", (tp_size,))
+tensor_parallel(model, device_mesh)
+```
+
+An end-to-end example can be found in `python/sglang/bench_latency.py`.
+You can run it with the following command:
+```bash
+$ python3 -m sglang.bench_latency --correct \
+  --model meta-llama/Meta-Llama-3-8B \
+  --json-model-override-args '{"architectures": ["TorchNativeLlamaForCausalLM"]}' \
+  --tensor-parallel-size 2 \
+  --disable-cuda-graph
+```
+We will eanble CUDA Graph support soon.
+"""
+
 import types
 from typing import Any, Dict, Iterable, Optional, Tuple
 
@@ -24,7 +49,10 @@ import torch
 from torch import nn
 from torch.nn.parameter import Parameter
 from transformers import LlamaConfig
-from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
@@ -41,35 +69,45 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
+tp_size = get_tensor_model_parallel_world_size()
+tp_rank = get_tensor_model_parallel_rank()
+
 
 def gate_up_proj_weight_loader(
     self,
     param: Parameter,
     loaded_weight: torch.Tensor,
-    loaded_shard_id: Optional[int] = None,
+    loaded_shard_id: int,
 ):
-    if loaded_shard_id is None:
-        shard_offsets: List[Tuple[int, int, int]] = []
+    # shard_id: (shard_offset, shard_size)
+    gate_up_offsets = {}
+    current_shard_offset = 0
     for i, output_size in enumerate(self.output_sizes):
-            shard_offsets.append((i, current_shard_offset, output_size))
+        # Everything shrinks by tp_size if TP enabled
+        output_size = output_size // tp_size
+        gate_up_offsets[i] = (current_shard_offset, output_size)
         current_shard_offset += output_size
-        for shard_id, shard_offset, shard_size in shard_offsets:
-            loaded_weight_shard = loaded_weight.narrow(
-                output_dim, shard_offset, shard_size
-            )
-            self.weight_loader(param, loaded_weight_shard, shard_id)
-    else:
+    # Re-size the param to the size after TP
+    if current_shard_offset != param.shape[0]:
+        # The clone will free the original, full tensor
+        param.data = param.data.narrow(0, 0, current_shard_offset).clone()
+
+    # Now load gate or up
     assert loaded_shard_id < len(self.output_sizes)
     param_data = param.data
-        shard_size = loaded_weight.shape[0]
-        shard_offset = loaded_shard_id * shard_size
+    shard_offset, shard_size = gate_up_offsets[loaded_shard_id]
     param_data = param_data.narrow(0, shard_offset, shard_size)
+    loaded_weight = loaded_weight.narrow(0, tp_rank * shard_size, shard_size)
     assert param_data.shape == loaded_weight.shape
     param_data.copy_(loaded_weight)
-    return
 
 
 class LlamaMLP(nn.Module):
+    _tp_plan = {
+        "gate_up_proj": "Colwise_Sharded",
+        "down_proj": "Rowwise",
+    }
+
     def __init__(
         self,
         hidden_size: int,
@@ -104,62 +142,44 @@ class LlamaMLP(nn.Module):
         return x
 
 
-def _get_shard_offset_mapping(self, loaded_shard_id: str):
-    shard_offset_mapping = {
-        "q": 0,
-        "k": self.num_heads * self.head_size,
-        "v": (self.num_heads + self.num_kv_heads) * self.head_size,
-        "total": (self.num_heads + 2 * self.num_kv_heads) * self.head_size,
-    }
-    return shard_offset_mapping.get(loaded_shard_id)
-
-
-def _get_shard_size_mapping(self, loaded_shard_id: str):
-    shard_size_mapping = {
-        "q": self.num_heads * self.head_size,
-        "k": self.num_kv_heads * self.head_size,
-        "v": self.num_kv_heads * self.head_size,
-    }
-    return shard_size_mapping.get(loaded_shard_id)
-
-
 def qkv_proj_weight_loader(
     self,
     param: Parameter,
     loaded_weight: torch.Tensor,
-    loaded_shard_id: Optional[str] = None,
+    loaded_shard_id: str,
 ):
-    if loaded_shard_id is None:
-        shard_offsets = [
-            # (shard_id, shard_offset, shard_size)
-            ("q", 0, self.total_num_heads * self.head_size),
-            (
-                "k",
-                self.total_num_heads * self.head_size,
-                self.total_num_kv_heads * self.head_size,
+    num_heads = self.num_heads // tp_size
+    num_kv_heads = self.num_kv_heads // tp_size
+    # shard_id: (shard_offset, shard_size)
+    qkv_offsets = {
+        "q": (0, num_heads * self.head_size),
+        "k": (num_heads * self.head_size, num_kv_heads * self.head_size),
+        "v": (
+            (num_heads + num_kv_heads) * self.head_size,
+            num_kv_heads * self.head_size,
         ),
-            (
-                "v",
-                (self.total_num_heads + self.total_num_kv_heads) * self.head_size,
-                self.total_num_kv_heads * self.head_size,
-            ),
-        ]
-        for shard_id, shard_offset, shard_size in shard_offsets:
-            loaded_weight_shard = loaded_weight.narrow(
-                param.output_dim, shard_offset, shard_size
-            )
-            self.weight_loader(param, loaded_weight_shard, shard_id)
-    else:
-        shard_offset = self._get_shard_offset_mapping(loaded_shard_id)
-        shard_size = self._get_shard_size_mapping(loaded_shard_id)
+    }
+    total_size = qkv_offsets["v"][0] + qkv_offsets["v"][1]
+    # Re-size the param to the size after TP
+    if total_size != param.shape[0]:
+        # The clone will free the original, full tensor
+        param.data = param.data.narrow(0, 0, total_size).clone()
+
+    # Now load q, k or v
+    shard_offset, shard_size = qkv_offsets[loaded_shard_id]
     param_data = param.data
     param_data = param_data.narrow(0, shard_offset, shard_size)
+    loaded_weight = loaded_weight.narrow(0, tp_rank * shard_size, shard_size)
     assert param_data.shape == loaded_weight.shape
     param_data.copy_(loaded_weight)
-    return
 
 
 class LlamaAttention(nn.Module):
+    _tp_plan = {
+        "qkv_proj": "Colwise_Sharded",
+        "o_proj": "Rowwise",
+    }
+
     def __init__(
         self,
         config: LlamaConfig,
@@ -176,7 +196,6 @@ class LlamaAttention(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
-        tp_size = get_tensor_model_parallel_world_size()
         self.total_num_heads = num_heads
         assert self.total_num_heads % tp_size == 0
         self.num_heads = self.total_num_heads // tp_size
@@ -205,20 +224,12 @@ class LlamaAttention(nn.Module):
             (self.total_num_heads + 2 * self.total_num_kv_heads) * self.head_dim,
             bias=False,
         )
-        self.qkv_proj.total_num_heads = self.total_num_heads
         self.qkv_proj.head_size = self.head_dim
-        self.qkv_proj.total_num_kv_heads = self.total_num_kv_heads
         self.qkv_proj.num_heads = self.total_num_heads
         self.qkv_proj.num_kv_heads = self.total_num_kv_heads
         self.qkv_proj.weight_loader = types.MethodType(
             qkv_proj_weight_loader, self.qkv_proj
         )
-        self.qkv_proj._get_shard_offset_mapping = types.MethodType(
-            _get_shard_offset_mapping, self.qkv_proj
-        )
-        self.qkv_proj._get_shard_size_mapping = types.MethodType(
-            _get_shard_size_mapping, self.qkv_proj
-        )
         self.qkv_proj.weight.weight_loader = self.qkv_proj.weight_loader
         self.qkv_proj.weight.output_dim = 0
         self.o_proj = torch.nn.Linear(
@@ -385,6 +396,7 @@ class TorchNativeLlamaForCausalLM(nn.Module):
         self.config = config
         self.quant_config = quant_config
         self.torchao_config = global_server_args_dict["torchao_config"]
+        self.supports_torch_tp = True
         self.model = LlamaModel(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)

test/srt/test_torch_tp.py

+import unittest
+
+from sglang.test.test_utils import is_in_ci, run_bench_latency
+
+
+class TestTorchTP(unittest.TestCase):
+    def test_torch_native_llama(self):
+        output_throughput = run_bench_latency(
+            "meta-llama/Meta-Llama-3-8B",
+            [
+                "--tp",
+                "2",
+                "--json-model-override-args",
+                '{"architectures": ["TorchNativeLlamaForCausalLM"]}',
+                "--disable-cuda-graph",
+            ],
+        )
+
+        if is_in_ci():
+            assert output_throughput > 0, f"{output_throughput=}"
+
+
+if __name__ == "__main__":
+    unittest.main()