Use DTensor-based tensor parallel

ghstack-source-id: b55b264d20bd2c0054f7248435fd605a452e876b
Pull Request resolved: #180

Author: kwen2501

model.py

@@ -156,30 +156,22 @@ def __init__(self, config: ModelArgs):
         super().__init__()
         assert config.dim % config.n_head == 0
 
-        total_head_dim = (config.n_head + 2 * config.n_local_heads) * config.head_dim
-        # key, query, value projections for all heads, but in a batch
-        self.wqkv = nn.Linear(config.dim, total_head_dim, bias=False)
+        self.wq = nn.Linear(config.dim, config.n_head * config.head_dim, bias=False)
+        self.wk = nn.Linear(config.dim, config.n_local_heads * config.head_dim, bias=False)
+        self.wv = nn.Linear(config.dim, config.n_local_heads * config.head_dim, bias=False)
         self.wo = nn.Linear(config.dim, config.dim, bias=False)
         self.kv_cache = None
 
         self.n_head = config.n_head
         self.head_dim = config.head_dim
         self.n_local_heads = config.n_local_heads
         self.dim = config.dim
-        self._register_load_state_dict_pre_hook(self.load_hook)
-
-    def load_hook(self, state_dict, prefix, *args):
-        if prefix + "wq.weight" in state_dict:
-            wq = state_dict.pop(prefix + "wq.weight")
-            wk = state_dict.pop(prefix + "wk.weight")
-            wv = state_dict.pop(prefix + "wv.weight")
-            state_dict[prefix + "wqkv.weight"] = torch.cat([wq, wk, wv])
 
     def forward(self, x: Tensor, freqs_cis: Tensor, mask: Tensor, input_pos: Optional[Tensor] = None) -> Tensor:
         bsz, seqlen, _ = x.shape
 
         kv_size = self.n_local_heads * self.head_dim
-        q, k, v = self.wqkv(x).split([self.dim, kv_size, kv_size], dim=-1)
+        q, k, v = self.wq(x), self.wk(x), self.wv(x)
 
         q = q.view(bsz, seqlen, self.n_head, self.head_dim)
         k = k.view(bsz, seqlen, self.n_local_heads, self.head_dim)

scripts/convert_hf_checkpoint.py

@@ -83,7 +83,7 @@ def convert_hf_checkpoint(
         original_dir = checkpoint_dir / "original"
         pattern = re.compile(r"^consolidated\.\d{2}\.pth$")
         bin_files = {bin for bin in original_dir.iterdir() if pattern.match(bin.name)}
-        
+
 
     def permute(w, n_head):
         dim = config.dim
@@ -116,13 +116,8 @@ def permute(w, n_head):
             if "wq" in key:
                 q = final_result[key]
                 k = final_result[key.replace("wq", "wk")]
-                v = final_result[key.replace("wq", "wv")]
-                q = permute(q, config.n_head)
-                k = permute(k, config.n_local_heads)
-                final_result[key.replace("wq", "wqkv")] = torch.cat([q, k, v])
-                del final_result[key]
-                del final_result[key.replace("wq", "wk")]
-                del final_result[key.replace("wq", "wv")]
+                final_result[key] = permute(q, config.n_head)
+                final_result[key.replace("wq", "wk")] = permute(k, config.n_local_heads)
     else:
         final_result = merged_result
     print(f"Saving checkpoint to {checkpoint_dir / 'model.pth'}")

tp.py

@@ -4,10 +4,13 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the root directory of this source tree.
 import os
+from enum import Enum
 from typing import List, Optional
 
 import torch
 import torch.distributed as dist
+from torch.distributed import DeviceMesh
+from torch.distributed.tensor.parallel import ColwiseParallel, RowwiseParallel, parallelize_module
 from torch import nn
 if os.uname().sysname != "Darwin":
     from torch.distributed import _functional_collectives as funcol
@@ -16,7 +19,7 @@
     funcol = None
 
 from model import Attention, FeedForward, Transformer
-from quantize import WeightOnlyInt4Linear
+from quantize import WeightOnlyInt4Linear, WeightOnlyInt8Linear
 
 
 def _get_rank() -> int:
@@ -33,6 +36,12 @@ def local_break():
 def _get_world_size() -> int:
     return int(os.environ.get("LOCAL_WORLD_SIZE", "1"))
 
+global device_mesh
+
+def _get_tp_mesh():
+    # device_mesh has only TP dimension for now
+    return device_mesh
+
 def maybe_init_dist() -> Optional[int]:
     try:
         # provided by torchrun
@@ -48,86 +57,97 @@ def maybe_init_dist() -> Optional[int]:
 
     torch.cuda.set_device(rank)
     dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
+
+    global device_mesh
+    device_mesh = dist.init_device_mesh(
+        "cuda",
+        (world_size,),  # Only TP dimension for now
+    )
     return rank
 
+class TPMode(Enum):
+    MANUAL = 0
+    DTENSOR = 1
 
-def _apply_tp_linear(linear: nn.Linear, style: str, weight_splits: List[int] = []) -> None:
+def _apply_tp_linear(linear: nn.Linear, style: str) -> None:
     rank = _get_rank()
     world_size = _get_world_size()
+    tp_mesh = _get_tp_mesh()
 
     # Linear's weight matrix is transposed, and is of shape
     # (linear.out_features, linear.in_features)
     dim_lookup = {
-        "colwise": (0, "out_features"),
-        "rowwise": (1, "in_features")
+        "colwise": (0, "out_features", ColwiseParallel()),
+        "rowwise": (1, "in_features", RowwiseParallel()),
     }
     assert style in dim_lookup
-    shard_dim, size_attr = dim_lookup[style]
+    shard_dim, size_attr, tp_plan = dim_lookup[style]
 
     # ensure we can shard evenly
     assert getattr(linear, size_attr) % world_size == 0
     def shard(x, dim):
         assert x.size(dim=dim) % world_size == 0
         return torch.tensor_split(x, world_size, dim=dim)[rank]
 
-    def shard_qkv(qkv, dim, weight_splits):
-        q, k, v = qkv.split(weight_splits, dim=dim)
-        q = shard(q, dim)
-        k = shard(k, dim)
-        v = shard(v, dim)
-        return torch.cat((q,k,v), dim=dim)
-
-    # shard
-    if weight_splits:
-        # attention
-        assert len(weight_splits) == 3
-
-        if isinstance(linear, WeightOnlyInt4Linear):
-            sharded_weight = shard_qkv(linear.weight, shard_dim, [i//8 for i in weight_splits])
-            linear.scales_and_zeros = shard_qkv(linear.scales_and_zeros, 1 - shard_dim, weight_splits)
-        else:
-            sharded_weight = shard_qkv(linear.weight, shard_dim, weight_splits)
-        if hasattr(linear, "scales") and style == "colwise":
-            linear.scales = shard_qkv(linear.scales, 0, weight_splits)
-    else:
-        sharded_weight = shard(linear.weight, shard_dim)
-        if isinstance(linear, WeightOnlyInt4Linear):
+    def shard_scale(linear, shard_dim):
+        if hasattr(linear, "scales_and_zeros"):
             linear.scales_and_zeros = shard(linear.scales_and_zeros, 1 - shard_dim)
             if style == "rowwise":
                 assert linear.scales_and_zeros.shape[0] * 32 == sharded_weight.shape[1] * sharded_weight.shape[2] * sharded_weight.shape[3]
                 assert linear.scales_and_zeros.shape[1] == sharded_weight.shape[0] * 8
-        if hasattr(linear, "scales") and style == "colwise":
-            linear.scales = shard(linear.scales, 0)
+        elif hasattr(linear, "scale"):
+            if style == "colwise":
+                linear.scales = shard(linear.scales, 0)
+
+    # shard
+    tp_mode: TPMode
+    if isinstance(linear, (WeightOnlyInt4Linear, WeightOnlyInt8Linear)):
+        # TODO: DTensor doesn't have a way to distribute quantized tensor yet.
+        # Should revisit when that capability is added.
+        sharded_weight = shard(linear.weight, shard_dim)
+        linear.weight = nn.Parameter(sharded_weight, requires_grad=False)
+        shard_scale(linear, shard_dim)
+        tp_mode = TPMode.MANUAL
+    else:
+        # Use DTensor based TP
+        parallelize_module(linear, tp_mesh, tp_plan)
+        tp_mode = TPMode.DTENSOR
 
     # local_break()
-    linear.weight = nn.Parameter(sharded_weight, requires_grad=False)
     setattr(linear, size_attr, getattr(linear, size_attr) // world_size)
 
     # shape info should still be synced
     # assert linear.weight.shape == (linear.out_features, linear.in_features)
+    return tp_mode
 
 
 def _apply_tp_ffn(mlp: FeedForward) -> None:
     assert hasattr(mlp, "w1")
     assert hasattr(mlp, "w3")
     assert hasattr(mlp, "w2")
 
-    _apply_tp_linear(mlp.w1, "colwise")
-    _apply_tp_linear(mlp.w3, "colwise")
-    _apply_tp_linear(mlp.w2, "rowwise")
+    tp_mode = _apply_tp_linear(mlp.w1, "colwise")
+    tp_mode = _apply_tp_linear(mlp.w3, "colwise")
+    tp_mode = _apply_tp_linear(mlp.w2, "rowwise")
 
-    world_size = _get_world_size()
-    mlp.register_forward_hook(lambda _module, _input, output: funcol.all_reduce(
-        output, "sum", list(range(world_size))))
+    if tp_mode == TPMode.MANUAL:
+        # In manual mode, we need to manually add an all-reduce at the end
+        world_size = _get_world_size()
+        mlp.register_forward_hook(lambda _module, _input, output: funcol.all_reduce(
+            output, "sum", list(range(world_size))))
 
 
 def _apply_tp_attn(attn: Attention) -> None:
-    assert hasattr(attn, "wqkv")
+    assert hasattr(attn, "wq")
+    assert hasattr(attn, "wk")
+    assert hasattr(attn, "wv")
     assert hasattr(attn, "wo")
 
     kv_size = attn.n_local_heads * attn.head_dim
-    _apply_tp_linear(attn.wqkv, "colwise", [attn.dim, kv_size, kv_size])
-    _apply_tp_linear(attn.wo, "rowwise")
+    tp_mode = _apply_tp_linear(attn.wq, "colwise")
+    tp_mode = _apply_tp_linear(attn.wk, "colwise")
+    tp_mode = _apply_tp_linear(attn.wv, "colwise")
+    tp_mode = _apply_tp_linear(attn.wo, "rowwise")
 
     # overwrite
     world_size = _get_world_size()
@@ -136,8 +156,10 @@ def _apply_tp_attn(attn: Attention) -> None:
     attn.head_dim = attn.dim // attn.n_head
     attn.n_local_heads = attn.n_local_heads // world_size
 
-    attn.register_forward_hook(lambda _module, _input, output: funcol.all_reduce(
-        output[0], "sum", list(range(world_size))))
+    if tp_mode == TPMode.MANUAL:
+        # In manual mode, we need to manually add an all-reduce at the end
+        attn.register_forward_hook(lambda _module, _input, output: funcol.all_reduce(
+            output[0], "sum", list(range(world_size))))
 
 
 def _apply_tp_Transformer(Transformer: Transformer) -> None: