Add rope fp32 (#43)

* Log model

* Add flag for rope outer in fp32

---------

Co-authored-by: Srini Iyer <sviyer@meta.com>

bytelatent/base_transformer.py

@@ -45,6 +45,7 @@ class BaseTransformerArgs(BaseModel):
     norm_eps: float = 1e-5
 
     rope_theta: float = 10000.0
+    rope_use_fp32_in_outer_product: bool = False
 
     init_base_std: float | None = None
     init_std_factor: InitStdFactor = InitStdFactor.DISABLED
@@ -78,7 +79,12 @@ def repeat_kv(x: torch.Tensor, n_rep: int, dim: int) -> torch.Tensor:
     )
 
 
-def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
+def precompute_freqs_cis(
+    dim: int,
+    end: int,
+    theta: float = 10000.0,
+    rope_use_fp32_in_outer_product: bool = False,
+):
     """
     Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
 
@@ -96,6 +102,9 @@ def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
     """
     freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
     t = torch.arange(end, device=freqs.device)
+    if rope_use_fp32_in_outer_product:
+        t = t.to(torch.float32)
+
     freqs = torch.outer(t, freqs).float()
 
     cos, sin = freqs.cos(), freqs.sin()
@@ -232,22 +241,37 @@ class RotaryEmbedding(torch.nn.Module):
     RotaryEmbedding Module
     """
 
-    def __init__(self, theta: float, head_dim: int, max_seqlen: int = 1024):
+    def __init__(
+        self,
+        theta: float,
+        head_dim: int,
+        max_seqlen: int = 1024,
+        rope_use_fp32_in_outer_product: bool = False,
+    ):
         super().__init__()
 
         self.theta = theta
         self.head_dim = head_dim
         self.max_seqlen = max_seqlen
+        self.rope_use_fp32_in_outer_product = rope_use_fp32_in_outer_product
 
         self.register_buffer(
             "freqs_cis",
-            precompute_freqs_cis(dim=head_dim, end=max_seqlen, theta=theta),
+            precompute_freqs_cis(
+                dim=head_dim,
+                end=max_seqlen,
+                theta=theta,
+                rope_use_fp32_in_outer_product=self.rope_use_fp32_in_outer_product,
+            ),
             persistent=False,
         )
 
     def reset_parameters(self):
         self.freqs_cis[...] = precompute_freqs_cis(
-            dim=self.head_dim, end=self.max_seqlen, theta=self.theta
+            dim=self.head_dim,
+            end=self.max_seqlen,
+            theta=self.theta,
+            rope_use_fp32_in_outer_product=self.rope_use_fp32_in_outer_product,
         )
 
     def forward(
@@ -577,6 +601,7 @@ class BaseTransformer(nn.Module):
             theta=args.rope_theta,
             head_dim=args.head_dim or args.dim // args.n_heads,
             max_seqlen=args.max_seqlen,
+            rope_use_fp32_in_outer_product=args.rope_use_fp32_in_outer_product,
         )
         self.eos_id = args.eos_id
 

bytelatent/model/blt.py

@@ -414,7 +414,7 @@ class ByteLatentTransformerArgs(BaseTransformerArgs):
     patch_in_forward: bool = False
 
     # Architecture and dimensions
-    dim_token: int = 256
+    dim_token: int | None = None
     dim_global: int = 512
     dim_local_decoder: int = 512
     dim_local_encoder: int = 512
@@ -523,10 +523,6 @@ class ByteLatentTransformerArgs(BaseTransformerArgs):
     use_fsdp: bool = True
     attn_to_keep: str = "all"
 
-    # RoPE parameters
-    rope_theta: float = 10000.0
-    rope_use_fp32_in_outer_product: bool = False
-
     # Parameter mixing
     pm_size: int = 0
 
@@ -619,6 +615,7 @@ def create_local_encoder(args: ByteLatentTransformerArgs) -> LocalEncoder:
         sliding_window=args.local_attention_window_len,
         use_rope=args.use_rope,
         rope_theta=args.rope_theta,
+        rope_use_fp32_in_outer_product=args.rope_use_fp32_in_outer_product,
         init_base_std=args.init_base_std,
         init_std_factor=args.init_std_factor,
         n_kv_heads=args.n_kv_heads,
@@ -661,6 +658,7 @@ def create_local_decoder(args: ByteLatentTransformerArgs) -> LocalDecoder:
         sliding_window=args.local_attention_window_len,
         use_rope=args.use_rope,
         rope_theta=args.rope_theta,
+        rope_use_fp32_in_outer_product=args.rope_use_fp32_in_outer_product,
         init_base_std=args.init_base_std,
         init_std_factor=args.init_std_factor,
         n_kv_heads=args.n_kv_heads,

bytelatent/model/local_models.py

@@ -86,6 +86,7 @@ class LocalModelBase(nn.Module):
                 theta=args.rope_theta,
                 head_dim=args.head_dim or args.dim // args.n_heads,
                 max_seqlen=args.max_seqlen,
+                rope_use_fp32_in_outer_product=args.rope_use_fp32_in_outer_product,
             )
             self.pos_embeddings = None
 

bytelatent/train.py

@@ -325,6 +325,7 @@ def train(args: TrainArgs):
 
         # log model size
 
+        logger.info(model)
         logger.info(f"Model size: {model_param_count:,} total parameters")
 
         gpu_memory_monitor = GPUMemoryMonitor("cuda")