fix: Mirror llama4 rope scaling fixes, small model simplify (#1917)

See:
- https://github.com/meta-llama/llama-models/pull/322
- https://github.com/meta-llama/llama-models/pull/320

llama_stack/models/llama/llama4/args.py

@@ -70,6 +70,9 @@ class ModelArgs(BaseModel):
     attention_chunk_size: Optional[int] = None
     rope_theta: float = 500000
     use_scaled_rope: bool = False
+    rope_scaling_factor: Optional[float] = None
+    rope_high_freq_factor: Optional[float] = None
+
     nope_layer_interval: Optional[int] = None  # No position encoding in every n layers
     use_qk_norm: bool = False
     # Set to True to enable inference-time temperature tuning (useful for very long context)
@@ -92,4 +95,14 @@ class ModelArgs(BaseModel):
             f"n_heads ({self.n_heads}) must be divisible by n_kv_heads ({self.n_kv_heads})"
         )
         assert self.dim % self.n_heads == 0, f"dim ({self.dim}) must be divisible by n_heads ({self.n_heads})"
+
+        if self.use_scaled_rope:
+            # NOTE: ideally these values should have come from params.json. However, we have
+            # shipped the models everywhere. Only Llama-4-Scout uses scaled rope and needs these
+            # specific values.
+            if self.rope_scaling_factor is None:
+                self.rope_scaling_factor = 16
+            if self.rope_high_freq_factor is None:
+                self.rope_high_freq_factor = 1
+
         return self

llama_stack/models/llama/llama4/model.py

@@ -23,37 +23,25 @@ from .ffn import FeedForward
 from .moe import MoE
 
 
+def rmsnorm(x, eps):
+    def _norm(y):
+        return y * torch.rsqrt(y.pow(2).mean(-1, keepdim=True) + eps)
+
+    return _norm(x.float()).type_as(x)
+
+
 class RMSNorm(torch.nn.Module):
     def __init__(self, dim: int, eps: float = 1e-6):
         super().__init__()
         self.eps = eps
         self.weight = nn.Parameter(torch.ones(dim))
 
-    def _norm(self, x):
-        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
-
     def forward(self, x):
-        output = self._norm(x.float()).type_as(x)
-        return output * self.weight
+        return rmsnorm(x, self.eps) * self.weight
 
 
-class L2Norm(torch.nn.Module):
-    def __init__(self, dim: int, eps: float = 1e-6):
-        super().__init__()
-        self.eps = eps
-
-    def _norm(self, x):
-        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
-
-    def forward(self, x):
-        return self._norm(x.float()).type_as(x)
-
-
-def apply_scaling(freqs: torch.Tensor):
-    # Values obtained from grid search
-    scale_factor = 8
+def apply_scaling(freqs: torch.Tensor, scale_factor: float, high_freq_factor: float):
     low_freq_factor = 1
-    high_freq_factor = 4
     old_context_len = 8192  # original llama3 length
 
     low_freq_wavelen = old_context_len / low_freq_factor
@@ -72,11 +60,18 @@ def apply_scaling(freqs: torch.Tensor):
     return torch.tensor(new_freqs, dtype=freqs.dtype, device=freqs.device)
 
 
-def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0, use_scaled: bool = False):
+def precompute_freqs_cis(
+    dim: int,
+    end: int,
+    theta: float,
+    use_scaled: bool,
+    scale_factor: float,
+    high_freq_factor: float,
+):
     freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
     t = torch.arange(end, device=freqs.device, dtype=torch.float32)
     if use_scaled:
-        freqs = apply_scaling(freqs)
+        freqs = apply_scaling(freqs, scale_factor, high_freq_factor)
     freqs = torch.outer(t, freqs)
     freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
     return freqs_cis
@@ -174,9 +169,7 @@ class Attention(nn.Module):
                 self.head_dim,
             )
         ).cuda()
-        self.qk_norm = None
-        if self.use_qk_norm:
-            self.qk_norm = L2Norm(args.norm_eps)
+        self.norm_eps = args.norm_eps
         self._register_load_state_dict_pre_hook(self.load_hook)
 
     def load_hook(
@@ -220,8 +213,8 @@ class Attention(nn.Module):
             xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
 
         if self.use_qk_norm:
-            xq = self.qk_norm(xq)
-            xk = self.qk_norm(xk)
+            xq = rmsnorm(xq, self.norm_eps)
+            xk = rmsnorm(xk, self.norm_eps)
 
         # We are applying temperature tuning (https://arxiv.org/abs/2501.19399) to NoPE layers, where
         # the inference-time temperature tuning function is customized to not affect short context
@@ -362,6 +355,8 @@ class Transformer(nn.Module):
             args.max_seq_len * 2,
             args.rope_theta,
             args.use_scaled_rope,
+            args.rope_scaling_factor,
+            args.rope_high_freq_factor,
         )
         vision_args = self.args.vision_args
         if vision_args: