Add in some missing grok specific model structure and constants

sharktank/sharktank/layers/paged_llama_attention_block.py

@@ -37,6 +37,7 @@ def __init__(
         head_count_kv: int,
         rms_epsilon: float,
         use_hf: bool = False,
+        use_grok: bool = False,
     ):
         super().__init__(theta)
         self.add_module(
@@ -53,6 +54,7 @@ def __init__(
         self.head_dim = head_dim
         self.head_count_kv = head_count_kv
         self.use_hf = use_hf
+        self.use_grok = use_grok
 
     def forward(
         self,
@@ -141,6 +143,8 @@ def repeat_kv(x: torch.Tensor) -> torch.Tensor:
 
         # Flash attention.
         attn_weights = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
+        if self.use_grok:
+            attn_weights = 30 * torch.tanh(attn_weights * (0.08838834764831845 / 30.0))
         self.assert_not_nan(attn_weights)
 
         # Apply attention mask.

sharktank/sharktank/models/grok/grok.py

@@ -147,6 +147,8 @@ def __init__(self, theta: Theta, config: LlamaModelConfig):
                     head_dim=hp.attn_head_dim,
                     head_count_kv=hp.attention_head_count_kv,
                     rms_epsilon=hp.attention_layer_norm_rms_epsilon,
+                    use_hf=True,
+                    use_grok=True,
                 )
             )
             self.attn_blocks.append(
@@ -250,6 +252,7 @@ def decode(
         )
 
         h = self.token_embedding(tokens)
+        h *= 78.38367176906169
         self.trace_tensor("mixtral.token_embedding", h)
 
         # Iterate over attention blocks.
@@ -278,4 +281,5 @@ def decode(
 
         h = self.output_norm(h)
         logits = self.output_lm_head(h)
+        logits = logits * 0.5773502691896257
         return logits