updates

examples/llama/model.py

@@ -82,15 +82,17 @@ def __init__(
         super().__init__()
         self.eps = eps
         self.dtype = dtype
-        self.weight = ark.parameter([dim], dtype)
+        self.weight = ark.parameter([dim], ark.fp32)
 
     def _norm(self, x):
         # x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
         return ark.rmsnorm(x)
 
     def forward(self, x):
+        x = ark.cast(x, ark.fp32)
         output = self._norm(x)
-        return ark.mul(output, ark.reshape(self.weight, [1, 1, -1]))
+        output = ark.mul(output, ark.reshape(self.weight, [1, 1, -1]))
+        return ark.cast(output, self.dtype)
 
 
 class ColumnParallelLinear(ark.Module):
@@ -369,18 +371,16 @@ def forward(
         )
         if freqs_cis is not None:
             xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
-
-        # TODO: enable kv cache and mask later
+        # TODO: enable kv cache later
         keys = xk
         values = xv
         # (bs, n_local_heads, seqlen, head_dim)
         xq = ark.transpose(xq, [0, 2, 1, 3])
-        keys = ark.transpose(keys, [0, 2, 1, 3])
         values = ark.transpose(values, [0, 2, 1, 3])
 
         # (bs, n_local_heads, head_dim, seqlen)
-        keys_transpose = ark.transpose(keys, [0, 1, 3, 2])
-        scores = ark.matmul(xq, keys_transpose)
+        keys = ark.transpose(keys, [0, 2, 3, 1])
+        scores = ark.matmul(xq, keys)
         scores = ark.scale(scores, 1.0 / math.sqrt(self.head_dim))
 
         if mask is not None:
@@ -394,8 +394,7 @@ def forward(
         output = ark.reshape(
             output, [bsz, seqlen, self.head_dim * self.n_local_heads]
         )
-        output = self.wo(output)
-        return output
+        return self.wo(output)
 
 
 class TransformerBlock(ark.Module):

examples/llama/model_test.py

@@ -12,7 +12,7 @@
 import llama.model as model_pt
 import model as model_ark
 import numpy as np
-from typing import Any, Dict, List
+from typing import Dict, List
 from model import ModelArgs, ModelArgs7B, ModelArgs13B, ModelArgs70B