Add truncated llama style model init via reset parameters() (#54)

This PR adds the following:
1 - via reset parameters, a full layerwise init for the llama models
under /llama. This uses the total model depth as part of the init via:
self.weight_init_std = 0.02 / (2 * self.num_layers) ** 0.5

2 - The final output ffn (head) is init with sqrt of the dim of the
model itself and a slightly wider cutoff factor of 3.

3 - tangential change - updates run_llama_train.sh with updated MODEL
and MODEL_CONF params to allow for direct model control via the sh
script. (there was a MODEL already but it was incorrectly using that in
place of MODEL_CONF...though we should update this as it's not
intuitive).

4 - made the debugmodel default to 2 layers as an improved debug check.

5 - added a 1B and 40B for additional testing configs. I can't currently
run 70B on my H100 due to OOM, but can run 40B.

Testing:
Verified proper init and training with 7B, 13B and ~40B:

<img width="1085" alt="Screenshot 2024-02-11 at 10 39 12 PM"
src="https://github.com/pytorch-labs/torchtrain/assets/46302957/049037ed-63a4-4ab0-bebc-f297857aab72">

[ghstack-poisoned]

Author: Andrew Gu

run_llama_train.sh

@@ -8,7 +8,8 @@ TRAINER_DIR=${1:-/home/$USER/local/torchtrain}
 # e.g.
 # LOG_RANK=0,1 NGPU=4 SP=2 ./run_llama_train.sh
 
-MODEL=${MODEL:-"debugmodel"}
+MODEL=${MODEL:-"llama"}
+MODEL_CONF=${MODEL_CONF:-"debugmodel"}
 NGPU=${NGPU:-"8"}
 PP=${PP:-"1"}
 SP=${SP:-"1"}
@@ -24,6 +25,8 @@ CHECKPOINT_INTERVAL=${CHECKPOINT_INTERVAL:-5}
 
 torchrun --nproc_per_node=${NGPU} \
 --local-ranks-filter ${LOG_RANK} --role rank --tee 3 \
-train.py --steps 10 --compile \
---pp_degree ${PP} --sp_degree ${SP} --dp_degree ${DP}
+train.py --steps 10 \
+--model ${MODEL} --model_conf ${MODEL_CONF} \
+--pp_degree ${PP} --sp_degree ${SP} --dp_degree ${DP} \
+--compile
 --checkpoint-folder=${CHECKPOINT_FOLDER} --checkpoint-interval=${CHECKPOINT_INTERVAL}

torchtrain/models/llama/__init__.py

@@ -6,9 +6,11 @@
 __all__ = ["Transformer"]
 
 llama_configs = {
-    "debugmodel": ModelArgs(dim=256, n_layers=1, n_heads=16),
+    "debugmodel": ModelArgs(dim=256, n_layers=2, n_heads=16),
+    "1B": ModelArgs(dim=1024, n_layers=16, n_heads=8),
     "7B": ModelArgs(dim=4096, n_layers=32, n_heads=32),
     "13B": ModelArgs(dim=5120, n_layers=40, n_heads=40),
+    "40B": ModelArgs(dim=5120, n_layers=80, n_heads=40),
     "70B": ModelArgs(
         dim=8192,
         n_layers=80,

torchtrain/models/llama/model.py

@@ -8,6 +8,8 @@
 import torch.nn.functional as F
 from torch import nn
 
+from torchtrain.logging_utils import rank0_log
+
 
 @dataclass
 class ModelArgs:
@@ -183,7 +185,6 @@ class Attention(nn.Module):
     """
 
     def __init__(self, model_args: ModelArgs):
-
         super().__init__()
         self.n_heads = model_args.n_heads
         self.n_kv_heads = (
@@ -203,6 +204,20 @@ def __init__(self, model_args: ModelArgs):
             model_args.n_heads * self.head_dim, model_args.dim, bias=False
         )
 
+    def reset_parameters(self, init_std):
+        for item in (self.wq, self.wk, self.wv):
+            nn.init.trunc_normal_(
+                item.weight,
+                mean=0.0,
+                std=0.02,
+            )
+
+        nn.init.trunc_normal_(
+            self.wo.weight,
+            mean=0.0,
+            std=init_std,
+        )
+
     def forward(
         self,
         x: torch.Tensor,
@@ -277,7 +292,6 @@ def __init__(
         multiple_of: int,
         ffn_dim_multiplier: Optional[float],
     ):
-
         super().__init__()
         hidden_dim = int(2 * hidden_dim / 3)
         # custom dim factor multiplier
@@ -292,6 +306,20 @@ def __init__(
     def forward(self, x):
         return self.w2(F.silu(self.w1(x)) * self.w3(x))
 
+    def reset_parameters(self, init_std):
+        nn.init.trunc_normal_(
+            self.w1.weight,
+            mean=0.0,
+            std=0.02,
+        )
+
+        for item in (self.w2, self.w3):
+            nn.init.trunc_normal_(
+                item.weight,
+                mean=0.0,
+                std=init_std,
+            )
+
 
 class RotaryEmbedding(nn.Module):
     """
@@ -350,7 +378,6 @@ class TransformerBlock(nn.Module):
     """
 
     def __init__(self, layer_id: int, model_args: ModelArgs):
-
         super().__init__()
         self.n_heads = model_args.n_heads
         self.dim = model_args.dim
@@ -362,8 +389,10 @@ def __init__(self, layer_id: int, model_args: ModelArgs):
             ffn_dim_multiplier=model_args.ffn_dim_multiplier,
         )
         self.layer_id = layer_id
+        self.num_layers = model_args.n_layers
         self.attention_norm = RMSNorm(model_args.dim, eps=model_args.norm_eps)
         self.ffn_norm = RMSNorm(model_args.dim, eps=model_args.norm_eps)
+        self.weight_init_std = 0.02 / (2 * self.num_layers) ** 0.5
 
     def forward(
         self,
@@ -385,6 +414,14 @@ def forward(
         out = h + self.feed_forward(self.ffn_norm(h))
         return out
 
+    def reset_parameters(self):
+        """reset params and norms for entire block"""
+        self.attention_norm.reset_parameters()
+        self.ffn_norm.reset_parameters()
+
+        self.attention.reset_parameters(self.weight_init_std)
+        self.feed_forward.reset_parameters(self.weight_init_std)
+
 
 class Transformer(nn.Module):
     """
@@ -406,11 +443,11 @@ class Transformer(nn.Module):
     """
 
     def __init__(self, model_args: ModelArgs):
-
         super().__init__()
         self.model_args = model_args
         self.vocab_size = model_args.vocab_size
         self.n_layers = model_args.n_layers
+        self.model_dim = model_args.dim
 
         self.embeddings = RotaryEmbedding(model_args)
 
@@ -421,6 +458,27 @@ def __init__(self, model_args: ModelArgs):
         self.norm = RMSNorm(model_args.dim, eps=model_args.norm_eps)
         self.output = nn.Linear(model_args.dim, model_args.vocab_size, bias=False)
 
+        # init model weights
+        self.reset_parameters()
+        rank0_log(f"Model built with: {self.model_args}")
+
+    def reset_parameters(
+        self,
+    ):
+        for layer in self.layers:
+            layer.reset_parameters()
+        self.norm.reset_parameters()
+        final_out_std = self.model_dim**-0.5
+        cutoff_factor = 3
+        nn.init.trunc_normal_(
+            self.output.weight,
+            mean=0.0,
+            std=final_out_std,
+            a=-cutoff_factor * final_out_std,
+            b=cutoff_factor * final_out_std,
+        )
+        rank0_log("Model fully initialized via reset_params")
+
     def forward(self, tokens: torch.Tensor):
         """
         Perform a forward pass through the Transformer model.

train.py

@@ -78,6 +78,7 @@ def main(args):
     world_mesh = parallel_dims.build_mesh(device_type="cuda")
 
     model_name = args.model
+    rank0_log(f"Building {model_name}")
     # build tokenizer
     tokenizer_type = model_name_to_tokenizer[model_name]
     tokenizer = create_tokenizer(tokenizer_type, args.tokenizer_path)
@@ -222,7 +223,7 @@ def main(args):
     parser.add_argument(
         "--warmup_pct",
         type=float,
-        default=0.10,
+        default=0.20,
         help="percentage of total training steps to use for warmup",
     )
     parser.add_argument(