Make models amenable to scan

README.md

@@ -169,77 +169,58 @@ Finally, each model may also provide a GPU "original" version that illustrates
 and attributes where this model code came from, if any. This also helps to
 show case what changes we have done to make it performant on TPU. The original
 version is not expected to be run.
+
 ## Contributing
 
 Contributions are welcome! Please feel free to submit a pull request.
 
 When developing, use `pip install -e '.[dev]'` to install dev dependencies such
 as linter and formatter.
 
-### How to run tests:
+### How to run tests
 
 ```sh
 pytest
 ```
 
-### How to run some of the tests, and re-run them whenever you change a file:
+### How to run some of the tests, and re-run them whenever you change a file
 
 ```sh
 tp -i test ... # replace with path to tests/directories
 ```
 
+### How to format
 
-### How to run HuggingFace transformer models
-Torchprime supports run with huggingface models by taking advantage of `tp run`.
-To use huggingface models, you can clone
-[huggingface/transformers](https://github.com/huggingface/transformers) under
-torchprime and name it as `local_transformers`. This allows you to pick any
-branch or make code modifications in transformers for experiment:
-```
-git clone https://github.com/huggingface/transformers.git local_transformers
-```
-If huggingface transformer doesn't exist, torchprime will automatically clone
-the repo and build the docker for experiment. To switch to huggingface models,
-add flag `--use-hf` to `tp run` command:
+```sh
+ruff format
 ```
-tp run --use-hf torchprime/hf_models/train.py
+
+### How to lint
+
+```sh
+ruff check [--fix]
 ```
 
+You can install a Ruff VSCode plugin to check errors and format files from
+the editor.
+
 ### How to run inside the docker container locally
+
 You can also run locally without XPK with docker. When running inside the docker
 container, it will use the same dependencies and build process as used in the
 XPK approach, improving the hermeticity and reliability.
-```
+
+```sh
 tp docker-run torchprime/torch_xla_models/train.py
 ```
-This will run the TorchPrime docker image locally. You can also add `--use-hf`
-to run HuggingFace model locally.
-```
-tp docker-run --use-hf torchprime/hf_models/train.py
-```
 
-### How to run locally without XPK:
-```
-tp dbrun torchprime/torch_xla_models/train.py
-```
 This will run the TorchPrime docker image locally. You can also add `--use-hf`
 to run HuggingFace model locally.
 
-### How to format:
-
 ```sh
-ruff format
-```
-
-### How to lint:
-
-```sh
-ruff check [--fix]
+tp docker-run --use-hf torchprime/hf_models/train.py
 ```
 
-You can install a Ruff VSCode plugin to check errors and format files from
-the editor.
-
 ## Run distributed training with local torch/torch_xla wheel
 
 Torchprime supports running with user specified torch and torch_xla wheels placed

torchprime/layers/sequential.py

@@ -0,0 +1,41 @@
+from typing import Any
+
+import torch.nn as nn
+
+PyTree = Any
+
+
+class HomogeneousSequential(nn.Sequential):
+  """
+  HomogenousSequential is a sequential container that requires all child modules
+  to be of the same type and have matching input/output shapes. In turn, it may be
+  compiled with the `scan` higher order operator to save compile time.
+  """
+
+  repeated_layer: type
+  """The type of the layer being looped over."""
+
+  def __init__(self, *args: nn.Module) -> None:
+    super().__init__(*args)
+    types = set(type(module) for module in args)
+    assert len(types) == 1, f"All modules must be of the same type. Got {types}"
+    self.repeated_layer = types.pop()
+
+  def forward(self, *input, **broadcasted_inputs: PyTree):
+    """
+    Much like `torch.nn.Sequential`, this takes `input` and forwards it to the
+    first module it contains. It then "chains" outputs to inputs sequentially for
+    each subsequent module, finally returning the output of the last module.
+    Different from `torch.nn.Sequential`, you may specify `broadcasted_inputs` via
+    keyword arguments. The same keyword arguments will be passed to every layer
+    without changes (i.e. "broadcasted").
+    """
+    for module in self:
+      input = module(*splat(input), **broadcasted_inputs)
+    return input
+
+
+def splat(input):
+  if not isinstance(input, list | tuple):
+    input = (input,)
+  return input

torchprime/torch_xla_models/llama/model.py

@@ -28,7 +28,9 @@
 from transformers.activations import ACT2FN
 from transformers.utils import logging
 
+from torchprime.layers.sequential import HomogeneousSequential
 from torchprime.rope.rope import RopeScaling, llama3_rope_frequencies
+from torchprime.torch_xla_models import offloading
 from torchprime.torch_xla_models.loss import cross_entropy_loss
 
 logger = logging.get_logger(__name__)
@@ -328,15 +330,20 @@ def forward(
     self,
     hidden_states: torch.Tensor,
     attention_mask: torch.Tensor | None = None,
-    position_ids: torch.LongTensor | None = None,
-  ) -> torch.FloatTensor:
+    position_ids: torch.Tensor | None = None,
+  ) -> torch.Tensor:
     """
     Args:
         hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
         attention_mask (`torch.FloatTensor`, *optional*):
             attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
             query_sequence_length, key_sequence_length)` if default attention is used.
     """
+    # This gives the `hidden_states` tensor a name so that we can layer specify
+    # to offload this tensor to host RAM to save memory. This is not a standard
+    # torch API because there is no such feature in PyTorch. Instead, the name
+    # becomes node metadata during FX graph capture.
+    hidden_states = offloading.offload_name(hidden_states, "decoder_input")
 
     residual = hidden_states
 
@@ -370,10 +377,12 @@ class LlamaModel(nn.Module):
   def __init__(self, config: DictConfig):
     super().__init__()
     self.vocab_size = config.vocab_size
-
     self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
-    self.layers = nn.ModuleList(
-      [
+
+    # `HomogeneousSequential` is similar to `nn.Sequential` but can be compiled with
+    # `scan` described in https://pytorch.org/xla/release/r2.6/features/scan.html.
+    self.layers = HomogeneousSequential(
+      *[
         LlamaDecoderLayer(config, layer_idx)
         for layer_idx in range(config.num_hidden_layers)
       ]
@@ -385,15 +394,19 @@ def forward(
     self,
     input_ids: torch.LongTensor,
     attention_mask: torch.FloatTensor | None = None,
-  ) -> torch.FloatTensor:
+  ) -> torch.Tensor:
+    # convert input ids to embeddings
     inputs_embeds = self.embed_tokens(input_ids)
 
-    position_ids = torch.arange(
-      inputs_embeds.shape[1], device=inputs_embeds.device
-    ).unsqueeze(0)
-
-    # Create a causal mask without calling the current method
     seq_length = inputs_embeds.size(1)
+
+    # TODO(https://github.com/pytorch/xla/issues/8783): Pass position_ids as `long()`
+    # when `scan` can take non-differentiable inputs.
+    position_ids = (
+      torch.arange(seq_length, device=inputs_embeds.device).unsqueeze(0).float()
+    )
+
+    # Create a causal attention mask
     causal_mask = torch.triu(
       torch.full((seq_length, seq_length), float("-inf"), device=inputs_embeds.device),
       diagonal=1,
@@ -403,14 +416,10 @@ def forward(
     if attention_mask is not None:
       causal_mask = causal_mask * attention_mask[:, None, None, :]
 
-    # embed positions
-    hidden_states = inputs_embeds
-
     # decoder layers
-    for decoder_layer in self.layers:
-      hidden_states = decoder_layer(
-        hidden_states, attention_mask=causal_mask, position_ids=position_ids
-      )
+    hidden_states = self.layers(
+      inputs_embeds, attention_mask=causal_mask, position_ids=position_ids
+    )
 
     hidden_states = self.norm(hidden_states)
     return hidden_states

torchprime/torch_xla_models/mixtral/model.py

@@ -31,6 +31,7 @@
 from torch import nn
 from torch.nn import init
 
+from torchprime.layers.sequential import HomogeneousSequential
 from torchprime.torch_xla_models.loss import cross_entropy_loss
 from torchprime.torch_xla_models.topology import get_num_slices
 
@@ -129,8 +130,8 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
   Returns:
       `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
   """
-  cos = cos[position_ids].unsqueeze(unsqueeze_dim)
-  sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+  cos = cos[position_ids.long()].unsqueeze(unsqueeze_dim)
+  sin = sin[position_ids.long()].unsqueeze(unsqueeze_dim)
   q_embed = (q * cos) + (rotate_half(q) * sin)
   k_embed = (k * cos) + (rotate_half(k) * sin)
   return q_embed, k_embed
@@ -204,7 +205,7 @@ def forward(
     self,
     hidden_states: torch.Tensor,
     attention_mask: torch.Tensor | None = None,
-    position_ids: torch.LongTensor | None = None,
+    position_ids: torch.Tensor | None = None,
   ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
     bsz, q_len, _ = hidden_states.size()
 
@@ -816,7 +817,9 @@ def load_balance_loss(self, top_k_indices, logits):
     return loss
 
   @xp.trace_me("MixtralMoeBlock")
-  def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+  def forward(
+    self, hidden_states: torch.Tensor
+  ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     batch_size, sequence_length, hidden_dim = hidden_states.shape
     hidden_states = hidden_states.view(-1, hidden_dim)
     # router_logits: (batch * sequence_length, n_experts)
@@ -851,6 +854,7 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
       case "dropping":
         hidden_states = hidden_states.view(batch_size, sequence_length, hidden_dim)
         mesh = xs.get_global_mesh()
+        assert mesh is not None
         selected_experts = selected_experts.view(
           batch_size, sequence_length, self.top_k
         )
@@ -895,7 +899,11 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         final_hidden_states = final_hidden_states.reshape(
           batch_size, sequence_length, hidden_dim
         )
-    return final_hidden_states, router_logits, loss
+      case _:
+        raise NotImplementedError(
+          f"Unsupported moe implementation {self.moe_implementation}"
+        )
+    return final_hidden_states, router_logits, torch.tensor(loss)
 
 
 class MixtralDecoderLayer(nn.Module):
@@ -915,9 +923,10 @@ def __init__(self, config: DictConfig, layer_idx: int):
   def forward(
     self,
     hidden_states: torch.Tensor,
+    cumulative_loss: torch.Tensor,
     attention_mask: torch.Tensor | None = None,
-    position_ids: torch.LongTensor | None = None,
-  ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+    position_ids: torch.Tensor | None = None,
+  ) -> tuple[torch.Tensor, torch.Tensor]:
     residual = hidden_states
 
     hidden_states = self.input_layernorm(hidden_states)
@@ -936,7 +945,7 @@ def forward(
     hidden_states, router_logits, loss = self.block_sparse_moe(hidden_states)
     hidden_states = residual + hidden_states
 
-    outputs = (hidden_states, loss)
+    outputs = (hidden_states, cumulative_loss + loss)
     return outputs
 
 
@@ -954,8 +963,11 @@ def __init__(self, config: DictConfig):
     self.config = config
     self.vocab_size = config.vocab_size
     self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
-    self.layers = nn.ModuleList(
-      [
+
+    # `HomogeneousSequential` is similar to `nn.Sequential` but can be compiled with
+    # `scan` described in https://pytorch.org/xla/release/r2.6/features/scan.html.
+    self.layers = HomogeneousSequential(
+      *[
         MixtralDecoderLayer(config, layer_idx)
         for layer_idx in range(config.num_hidden_layers)
       ]
@@ -978,12 +990,14 @@ def _init_weights(self, module):
 
   @xp.trace_me("MixtralModel")
   def forward(
-    self, input_ids: torch.LongTensor = None, attention_mask: torch.Tensor | None = None
+    self, input_ids: torch.LongTensor, attention_mask: torch.Tensor | None = None
   ) -> tuple:
     batch_size, seq_length = input_ids.shape
 
     device = input_ids.device
-    position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
+    # TODO(https://github.com/pytorch/xla/issues/8783): Pass position_ids as `long()`
+    # when `scan` can take non-differentiable inputs.
+    position_ids = torch.arange(seq_length, dtype=torch.long, device=device).float()
     position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
 
     inputs_embeds = self.embed_tokens(input_ids)
@@ -999,22 +1013,18 @@ def forward(
 
     hidden_states = inputs_embeds
 
-    total_loss = 0.0
-    for decoder_layer in self.layers:
-      layer_outputs = decoder_layer(
-        hidden_states,
-        attention_mask=causal_mask,
-        position_ids=position_ids,
-      )
-
-      hidden_states = layer_outputs[0]
-      load_balance_loss = layer_outputs[-1]
-      total_loss += load_balance_loss
+    load_balance_loss = torch.tensor(0.0, device=device)
+    hidden_states, load_balance_loss = self.layers(
+      hidden_states,
+      load_balance_loss,
+      attention_mask=causal_mask,
+      position_ids=position_ids,
+    )
 
-    total_loss = total_loss / len(self.layers)
+    load_balance_loss = load_balance_loss / len(self.layers)
 
     hidden_states = self.norm(hidden_states)
-    return (hidden_states, total_loss)
+    return (hidden_states, load_balance_loss)
 
 
 class MixtralForCausalLM(nn.Module):