not working, incorrect handling with cache probably

llms/mlx_lm/models/cache.py

@@ -341,21 +341,20 @@ class MambaCache(_BaseCache):
 
 
 class Mamba2Cache(_BaseCache):
-    """Cache for Mamba model inference containing conv cache and SSM state."""
+    conv_states: Optional[mx.array] = None
-    conv_cache: Optional[mx.array] = None
+    ssm_states: Optional[mx.array] = None
-    ssm_state: Optional[mx.array] = None
 
     def __getitem__(self, idx: int) -> Optional[mx.array]:
         if idx == 0:
-            return self.conv_cache
+            return self.conv_states
         elif idx == 1:
-            return self.ssm_state
+            return self.ssm_states
         raise IndexError("Cache index must be 0 or 1")
 
     def __setitem__(self, idx: int, value: Optional[mx.array]):
         if idx == 0:
-            self.conv_cache = value
+            self.conv_states = value
         elif idx == 1:
-            self.ssm_state = value
+            self.ssm_states = value
         else:
             raise IndexError("Cache index must be 0 or 1")

llms/mlx_lm/models/mamba2.py

@@ -5,7 +5,7 @@ import mlx.core as mx
 import mlx.nn as nn
 
 from .base import BaseModelArgs
-from .cache import MambaCache
+from .cache import Mamba2Cache
 
 @dataclass
 class ModelArgs(BaseModelArgs):
@@ -62,6 +62,7 @@ def silu(x):
     return x * mx.sigmoid(x)
 
 def ssd(x, A, B, C, chunk_size):
+    # Replace einsum operations with explicit reshape and matrix multiply
     batch, seqlen, nheads, dim = x.shape
     B = mx.expand_dims(B, axis=2)
     C = mx.expand_dims(C, axis=2)
@@ -73,9 +74,18 @@ def ssd(x, A, B, C, chunk_size):
         chunk = slice(i, min(i + chunk_size, seqlen))
         dA = mx.exp(mx.expand_dims(A[chunk], axis=0))
         
-        dBx = mx.einsum('blhp,bln->bhpn', x[:, chunk], B[:, chunk])
+        # Replace einsum with explicit operations
+        x_chunk = x[:, chunk]  # [batch, chunk_size, nheads, dim]
+        x_chunk = mx.transpose(x_chunk, [0, 2, 3, 1])  # [batch, nheads, dim, chunk_size]
+        B_chunk = B[:, chunk]  # [batch, chunk_size, state_size]
+        dBx = mx.matmul(x_chunk, B_chunk)  # [batch, nheads, dim, state_size]
+        
         state = state * mx.expand_dims(dA, axis=-1) + dBx
-        y = mx.einsum('bhpn,bln->blhp', state, C[:, chunk])
+        
+        # Replace einsum with explicit operations
+        C_chunk = C[:, chunk]  # [batch, chunk_size, state_size]
+        y = mx.matmul(state, mx.transpose(C_chunk, [0, 2, 1]))  # [batch, nheads, dim, chunk_size]
+        y = mx.transpose(y, [0, 3, 1, 2])  # [batch, chunk_size, nheads, dim]
         outputs.append(y)
     
     return mx.concatenate(outputs, axis=1), state
@@ -93,7 +103,7 @@ class DepthWiseConv1d(nn.Module):
         assert in_channels == out_channels, "In and out channels must be same for depthwise convolution"
         assert self.groups == in_channels, "Groups must be equal to in_channels for depthwise convolution"
         
-        # Initialize with shape (channels, 1, kernel_size) to match pretrained weights
+        # Weight shape: (channels, 1, kernel_size) to match pretrained weights
         self.weight = mx.random.normal((in_channels, 1, kernel_size))
         self.bias = mx.zeros((out_channels,)) if bias else None
 
@@ -101,56 +111,78 @@ class DepthWiseConv1d(nn.Module):
         B, L, C = x.shape
         K = self.kernel_size
 
+        # Validate input dimensions
+        assert C == self.in_channels, f"Input channels {C} doesn't match expected {self.in_channels}"
+        
         # Handle padding and caching
         if cache is not None:
-            conv_cache = cache[cache_idx]
+            conv_states = cache[cache_idx]
-            if conv_cache is not None:
+            if conv_states is not None:
-                x = mx.concatenate([conv_cache, x], axis=1)
+                # Validate cache shape
-                L = x.shape[1]  # Update L after concatenation
+                assert conv_states.shape[0] == B, "Cache batch size mismatch"
+                assert conv_states.shape[2] == C, "Cache channel count mismatch"
+                x = mx.concatenate([conv_states, x], axis=1)
+                L = x.shape[1]
         else:
+            # Add left padding of size (kernel_size - 1)
             pad_left = K - 1
             x = mx.pad(x, [(0, 0), (pad_left, 0), (0, 0)])
-            L = x.shape[1]  # Update L after padding
+            L = x.shape[1]
 
-        # Implement depthwise convolution manually for each channel
+        # Pre-allocate output array if possible
         outputs = []
+        
+        # Process each channel independently
         for c in range(C):
-            # Extract single channel and reshape for 1D convolution
+            # Extract and prepare channel data
             x_c = x[:, :, c]  # Shape: [B, L]
             x_c = mx.expand_dims(x_c, axis=1)  # Shape: [B, 1, L]
             
-            # Extract and ensure filter is 3D
+            # Prepare filter weights
-            w_c = self.weight[c]  # Shape: [1, kernel_size] or [1, 1, kernel_size]
+            w_c = self.weight[c]  # Get channel weights
+            # Ensure filter is 3D: [depth(1), in_channels(1), kernel_size]
             if w_c.ndim == 2:
-                w_c = mx.expand_dims(w_c, axis=0)  # Shape: [1, 1, kernel_size]
+                w_c = mx.expand_dims(w_c, axis=0)
             elif w_c.ndim == 1:
                 w_c = mx.expand_dims(mx.expand_dims(w_c, axis=0), axis=0)
             
-            # For inference mode (single token), adjust the input
+            # Handle inference mode (single token)
             if L < K:
-                # Pad input to match kernel size
                 pad_size = K - L
                 x_c = mx.pad(x_c, [(0, 0), (0, 0), (pad_size, 0)])
             
-            # Apply 1D convolution for this channel
+            # Apply 1D convolution
+            try:
                 y_c = mx.conv_general(
                     x_c,
                     w_c,
                     stride=1,
-                padding=0  # We've already handled padding
+                    padding=0  # Padding already handled
                 )
                 
                 if self.bias is not None:
                     y_c = y_c + self.bias[c]
                 
-            outputs.append(mx.squeeze(y_c, axis=1))  # Shape: [B, 1]
+                # Remove singleton dimension and add to outputs
+                outputs.append(mx.squeeze(y_c, axis=1))
                 
-        # Stack all channel outputs
+            except Exception as e:
+                raise RuntimeError(f"Convolution failed for channel {c}. Shapes: input={x_c.shape}, weight={w_c.shape}") from e
+
+        # Stack channel outputs along last dimension
         y = mx.stack(outputs, axis=-1)  # Shape: [B, L', C]
         
+        # Update cache if needed
         if cache is not None:
-            # Update cache with the most recent K-1 tokens
+            # Store last (kernel_size - 1) tokens or entire input if shorter
-            cache[cache_idx] = x[:, -(K-1):, :] if L >= K else x
+            new_cache = x[:, -(K-1):, :] if L >= K else x
+            cache[cache_idx] = new_cache
+
+            if new_cache.shape != cache[cache_idx].shape:
+                cache[cache_idx] = new_cache
+                print(f"Cache updated at index {cache_idx}")
+            else:
+                print(f"Skipping cache update at index {cache_idx}, shapes are identical.")
 
         return y
     
@@ -184,9 +216,10 @@ class Mamba2Block(nn.Module):
             layer_scale = math.sqrt(1.0 / args.num_hidden_layers)
             self.out_proj.weight = self.out_proj.weight * layer_scale
 
-    def __call__(self, u: mx.array, cache = None):
+    def __call__(self, x: mx.array, cache=None):
-        if cache is not None and self.args.use_cache:
+        # if cache is not None and self.args.use_cache:
-            return self.step(u, cache)
+        if cache is not None:
+            return self.step(x, cache)
 
         # Calculate sizes
         d_model = self.args.intermediate_size
@@ -197,7 +230,7 @@ class Mamba2Block(nn.Module):
         A = -mx.exp(self.A_log)
         
         # Project input
-        zxbcdt = self.in_proj(u)
+        zxbcdt = self.in_proj(x)
         
         # Correct splits for z, xBC, dt
         splits = [
@@ -262,13 +295,7 @@ class Mamba2Block(nn.Module):
 
         return y
 
-    def step(self, u: mx.array, cache: MambaCache):
+    def step(self, u: mx.array, cache):
-        """
-        Process single or multiple tokens while maintaining state.
-        Args:
-            u: Input tensor of shape (batch_size, seq_len, hidden_size)
-            cache: MambaCache object containing conv cache and ssm state
-        """
         batch_size = u.shape[0]
         seq_len = u.shape[1]
         outputs = []
@@ -295,17 +322,11 @@ class Mamba2Block(nn.Module):
             n_heads = self.args.num_heads
             d_head = self.args.head_dim
             
-            # Correct splits for z, xBC, dt
+            # Split projected input
-            splits = [
+            # conv_dim = d_model + 2 * d_state (this should match self.conv1d.in_channels)
-                d_model,                    # z size
+            z = zxbcdt[:, :, :d_model]
-                d_model + 2 * d_state,      # xBC size (delta, B, C)
+            xBC = zxbcdt[:, :, d_model:d_model + 2*d_state + d_model]  # Include the full conv dimension
-                n_heads                     # dt size
+            dt = zxbcdt[:, :, -(n_heads):]
-            ]
-            
-            # Split the projected input
-            z = zxbcdt[:, :, :splits[0]]
-            xBC = zxbcdt[:, :, splits[0]:splits[0] + splits[1]]
-            dt = zxbcdt[:, :, -splits[2]:]  # Take last n_heads elements
             
             # Process dt
             dt = mx.reshape(dt, (batch_size, n_heads))
@@ -316,25 +337,23 @@ class Mamba2Block(nn.Module):
             )
             dt = mx.maximum(dt, self.args.time_step_floor)
 
-            # Process convolution
+            # Process convolution with correct dimensions
             xBC = self.conv1d(xBC, cache=cache, cache_idx=0)
             xBC = silu(xBC)
 
-            # Split convolved xBC into x, B, C
+            # Split convolved xBC into x, B, C with correct dimensions
             x = xBC[:, :, :d_model]
             B = xBC[:, :, d_model:d_model + d_state]
             C = xBC[:, :, -d_state:]
             
-            # Reshape x into (batch, heads, dim)
+            # Reshape tensors for SSM computation
             x = mx.reshape(x, (batch_size, 1, n_heads, d_head))
             x = mx.squeeze(x, axis=1)  # (batch, heads, dim)
             
-            # Reshape B into (batch, heads, dim, state)
             B = mx.reshape(B, (batch_size, 1, d_state))
             B = mx.broadcast_to(B, (batch_size, n_heads, d_state))
             B = mx.expand_dims(B, axis=2)  # (batch, heads, 1, state)
             
-            # Reshape C for later use
             C = mx.reshape(C, (batch_size, 1, d_state))
             C = mx.broadcast_to(C, (batch_size, n_heads, d_state))
             C = mx.expand_dims(C, axis=3)  # (batch, heads, state, 1)
@@ -344,14 +363,11 @@ class Mamba2Block(nn.Module):
             dA = mx.exp(dt * mx.expand_dims(A, 0))
             dA = mx.expand_dims(mx.expand_dims(dA, -1), -1)  # (batch, heads, 1, 1)
 
-            # Prepare x for Bx computation
+            # Update state with proper shapes
             x = mx.expand_dims(x, axis=3)  # (batch, heads, dim, 1)
-            
-            # Compute dBx with proper broadcasting
             dBx = mx.matmul(x, B)  # (batch, heads, dim, state)
             
-            # Update state
+            ssm_state = cache[1]
-            ssm_state = cache[1]  # (batch, heads, dim, state)
             ssm_state = ssm_state * dA + dBx
             cache[1] = ssm_state
 
@@ -359,18 +375,16 @@ class Mamba2Block(nn.Module):
             y = mx.matmul(ssm_state, C)  # (batch, heads, dim, 1)
             y = mx.squeeze(y, axis=-1)  # (batch, heads, dim)
             
-            # Add skip connection with D
+            # Add skip connection
             y = y + x[:, :, :, 0] * mx.expand_dims(self.D, -1)
             
-            # Reshape to original dimensions
+            # Reshape and process output
             y = mx.reshape(y, (batch_size, 1, n_heads * d_head))
-
-            # Apply norm and output projection
             y = self.norm(y + z)
             y = self.out_proj(y)
 
             if self.args.residual_in_fp32:
-                y.astype(mx.float32)
+                y = y.astype(mx.float32)
 
             outputs.append(y)
 
@@ -428,8 +442,8 @@ class Model(nn.Module):
         print('ouput')
         return logits
 
-    def make_cache(self):
+    def make_cache(self, batch_size=1):
-        return [MambaCache() for _ in range(len(self.layers))]
+        return [Mamba2Cache(batch_size, self.args.num_heads, self.args.head_dim, self.args.state_size) for _ in range(len(self.layers))]
     
     def sanitize(self, weights):
         sanitized = {}