[Not for land] Added changes for GPT-2 perf

ghstack-source-id: 39b91d06c8c1c6398e58a7d8841c4432ba4532c7
Pull Request resolved: #533

torchtitan/models/llama/model.py

@@ -127,6 +127,60 @@ def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
     )
 
 
+class ChunkedCE(torch.autograd.Function):
+    """
+    Credit: https://github.com/Chillee
+    """
+
+    @staticmethod
+    def forward(ctx, _input, weight, target, compiled=True):
+        CHUNK_SIZE = 8192
+        inp_shape = _input.shape
+        _input = _input.view(-1, _input.shape[-1])
+        target = target.view(-1)
+
+        def compute_loss(input_chunk, weight, target):
+            logits = torch.mm(input_chunk, weight.t())
+            logits = logits.float()
+            loss = torch.nn.functional.cross_entropy(logits, target)
+            return loss
+
+        grad_weight = torch.zeros_like(weight)
+        grad_inputs = []
+        loss_acc = torch.zeros((), device=_input.device)
+
+        chunks = max(_input.shape[0] // CHUNK_SIZE, 1)
+
+        def accumulate_chunk(input_chunk, target_chunk):
+            (chunk_grad_input, chunk_grad_weight), chunk_loss = (
+                torch.func.grad_and_value(compute_loss, argnums=(0, 1))(
+                    input_chunk, weight, target_chunk
+                )
+            )
+            grad_weight.add_(chunk_grad_weight)
+            loss_acc.add_(chunk_loss)
+            return chunk_grad_input
+
+        if compiled:
+            accumulate_chunk = torch.compile(accumulate_chunk)
+
+        input_chunks = torch.chunk(_input, chunks=chunks, dim=0)
+        target_chunks = torch.chunk(target, chunks=chunks, dim=0)
+        for input_chunk, target_chunk in zip(input_chunks, target_chunks):
+            grad_inputs.append(accumulate_chunk(input_chunk, target_chunk))
+
+        ctx.save_for_backward(
+            (torch.cat(grad_inputs, dim=0) / chunks).view(inp_shape),
+            grad_weight / chunks,
+        )
+        return loss_acc / chunks
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        (grad_input, grad_weight) = ctx.saved_tensors
+        return (grad_input, grad_weight, None, None)
+
+
 class Attention(nn.Module):
     """
     Multi-head attention module.
@@ -421,7 +475,7 @@ def _precompute_freqs_cis(self) -> torch.Tensor:
             self.model_args.rope_theta,
         )
 
-    def forward(self, tokens: torch.Tensor):
+    def forward(self, tokens: torch.Tensor, labels: torch.Tensor):
         """
         Perform a forward pass through the Transformer model.
 
@@ -439,8 +493,9 @@ def forward(self, tokens: torch.Tensor):
             h = layer(h, self.freqs_cis)
 
         h = self.norm(h) if self.norm else h
-        output = self.output(h).float() if self.output else h
-        return output
+        if not self.output:
+            return h
+        return ChunkedCE.apply(h, self.output.weight, labels)
 
     @classmethod
     def from_model_args(cls, model_args: ModelArgs) -> "Transformer":

torchtitan/parallelisms/parallelize_llama.py

@@ -312,15 +312,15 @@ def apply_fsdp(
             # all-gathers, which can be expensive and non-overlapped
             reshard_after_forward = False
         else:
-            # As an optimization, do not reshard after forward for the last
-            # transformer block since FSDP would prefetch it immediately
-            reshard_after_forward = int(layer_id) < len(model.layers) - 1
+            # For small models (e.g. GPT-2), parameter memory is low, so there
+            # is no need to reshard after forward
+            reshard_after_forward = False
         fully_shard(
             transformer_block,
             **fsdp_config,
             reshard_after_forward=reshard_after_forward,
         )
-    fully_shard(model, **fsdp_config, reshard_after_forward=not pp_enabled)
+    fully_shard(model, **fsdp_config)
 
     logger.info("Applied FSDP to the model")
 

train.py

@@ -8,6 +8,7 @@
 import os
 import time
 from datetime import timedelta
+from typing import List
 
 import torch
 from torch.distributed.elastic.multiprocessing.errors import record
@@ -136,6 +137,12 @@ def main(job_config: JobConfig):
 
     # loss function to be shared by Pipeline Parallel and SPMD training
     def loss_fn(pred, labels):
+        if isinstance(pred, torch.Tensor):
+            pred_chunks = pred.chunk(token_chunked_cross_entropy_loss.num_chunks, dim=1)
+        else:
+            assert isinstance(pred, list)
+            pred_chunks = pred
+        return token_chunked_cross_entropy_loss(pred_chunks, labels)
         return torch.nn.functional.cross_entropy(
             pred.flatten(0, 1), labels.flatten(0, 1)
         )
@@ -288,11 +295,7 @@ def loss_fn(pred, labels):
             else:
                 # Non-PP forward / backward
                 with train_context():
-                    pred = model(input_ids)
-                    loss = loss_fn(pred, labels)
-                    # pred.shape=(bs, seq_len, vocab_size)
-                    # need to free to before bwd to avoid peaking memory
-                    del pred
+                    loss = model(input_ids, labels)
                     loss.backward()
 
             # clip gradients

train_configs/gpt2.toml

@@ -6,9 +6,9 @@ dump_folder = "./outputs"
 description = "GPT-2 training"
 
 [profiling]
-enable_profiling = false
+enable_profiling = true
 save_traces_folder = "profile_trace"
-profile_freq = 10
+profile_freq = 50
 
 [metrics]
 log_freq = 10
@@ -27,11 +27,11 @@ lr = 3e-4
 fused = true
 
 [training]
-batch_size = 16
+batch_size = 32
 seq_len = 8192
 warmup_steps = 200  # lr scheduler warm up
 max_norm = 1.0  # grad norm clipping
-steps = 100
+steps = 50
 data_parallel_degree = -1
 tensor_parallel_degree = 1
 compile = true