quick save

llms/mlx_lm/models/cache.py

@@ -338,3 +338,30 @@ class MambaCache(_BaseCache):
     @state.setter
     def state(self, v):
         self.cache = v
+
+
+class Mamba2Cache:
+    def __init__(self, num_layers):
+        self.conv_states = [None] * num_layers
+        self.ssm_states = [None] * num_layers
+        self.seqlen_offset = 0
+
+    def __getitem__(self, idx):
+        return (self.conv_states[idx], self.ssm_states[idx])
+
+    def __setitem__(self, idx, value):
+        self.conv_states[idx], self.ssm_states[idx] = value
+
+    @property
+    def state(self):
+        return {
+            'conv_states': self.conv_states,
+            'ssm_states': self.ssm_states,
+            'seqlen_offset': self.seqlen_offset
+        }
+
+    @state.setter
+    def state(self, v):
+        self.conv_states = v['conv_states']
+        self.ssm_states = v['ssm_states']
+        self.seqlen_offset = v['seqlen_offset']

llms/mlx_lm/models/mamba2 copy.py

@@ -6,41 +6,37 @@ from typing import Tuple, Union
 
 import mlx.core as mx
 import mlx.nn as nn
-
 from .base import BaseModelArgs
 
+# python -m mlx_lm.generate --model rokyang/mamba2-130m-hf  --prompt "hello how are you."
 
 @dataclass
 class ModelArgs(BaseModelArgs):
-    model_type: str = "mamba2"
+    num_heads: int
-    num_heads: int = 128
+    head_dim: int
-    head_dim: int = 64
+    vocab_size: int
-    vocab_size: int = 32768
+    hidden_size: int
-    hidden_size: int = 4096
+    state_size: int
-    state_size: int = 128
+    num_hidden_layers: int
-    num_hidden_layers: int = 64
+    layer_norm_epsilon: float
-    layer_norm_epsilon: float = 1e-5
+    expand: int
-    pad_token_id: int = 1
+    conv_kernel: int
-    bos_token_id: int = 0
+    n_groups: int
-    eos_token_id: int = 2
+    use_bias: bool
-    expand: int = 2
+    use_conv_bias: bool
-    conv_kernel: int = 4
+    initializer_range: float 
-    n_groups: int = 8
+    residual_in_fp32: bool
-    use_bias: bool = False
+    time_step_min: float
-    use_conv_bias: bool = True
+    time_step_max: float
-    hidden_act: str = "silu"
+    time_step_floor: float
-    initializer_range: float = 0.1
+    rescale_prenorm_residual: bool
-    residual_in_fp32: bool = True
+    use_cache: bool
-    time_step_rank: Union[int, str] = "auto"
+    rms_norm: bool
-    time_step_min: float = 0.001
+    chunk_size: int
-    time_step_max: float = 0.1
+    tie_word_embeddings: bool
-    time_step_floor: float = 1e-4
     time_step_limit: Tuple[float, float] = field(default_factory=lambda: (0.0, float("inf")))
-    rescale_prenorm_residual: bool = False
+    time_step_rank: Union[int, str] = "auto"
-    use_cache: bool = True
+    model_type: str = "mamba2"
-    rms_norm: bool = True
-    chunk_size: int = 256
-    tie_word_embeddings: bool = False
 
     def __post_init__(self):
         if not hasattr(self, "intermediate_size"):
@@ -79,15 +75,24 @@ class MambaRMSNormGated(nn.Module):
         hidden_states = hidden_states * mx.rsqrt(variance + self.variance_epsilon)
         return self.weight * hidden_states
 
+
 class DepthWiseConv1d(nn.Module):
-    def __init__(self, channels, kernel_size, bias=True, groups=1, padding=0):
+    def __init__(self, in_channels, out_channels, kernel_size, bias=True, groups=None, padding=0):
         super().__init__()
-        self.channels = channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
         self.kernel_size = kernel_size
         self.padding = padding
-        self.groups = groups
+        self.groups = groups if groups is not None else in_channels
-        self.weight = mx.random.normal((self.channels, kernel_size, 1))
+
-        self.bias = mx.zeros((channels,)) if bias else None
+        # Ensure in_channels and out_channels are the same for depthwise conv
+        assert in_channels == out_channels, "In and out channels must be the same for depthwise convolution"
+        # Ensure groups is equal to in_channels for depthwise conv
+        assert self.groups == in_channels, "Groups must be equal to in_channels for depthwise convolution"
+
+        # Initialize weight with shape (out_channels, kernel_size, 1)
+        self.weight = mx.random.normal((out_channels, kernel_size, 1))
+        self.bias = mx.zeros((out_channels,)) if bias else None
 
     def __call__(self, x, cache=None):
         B, L, C = x.shape
@@ -116,16 +121,17 @@ class Mamba2Mixer(nn.Module):
         self.hidden_size = args.hidden_size
         self.state_size = args.state_size
         self.num_heads = args.num_heads
-        self.head_dim = args.head_dim
+        self.head_dim = args.hidden_size // args.num_heads
         self.n_groups = args.n_groups
 
         self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.state_size
         self.conv1d = DepthWiseConv1d(
-            channels=self.conv_dim,
+            in_channels=self.conv_dim,
-            kernel_size=self.conv_kernel_size,
+            out_channels=self.conv_dim,
-            bias=self.args.use_conv_bias,
+            bias=args.use_conv_bias,
+            kernel_size=args.conv_kernel,
             groups=self.conv_dim,
-            padding=self.conv_kernel_size - 1,
+            padding=args.conv_kernel - 1
         )
 
         projection_size = self.intermediate_size + self.conv_dim + self.num_heads
@@ -135,33 +141,35 @@ class Mamba2Mixer(nn.Module):
             bias=args.use_bias
         )
 
-        self.act = nn.SiLU()
+        self.A_log = mx.zeros(self.num_heads)
-        self.dt_bias = mx.ones((self.num_heads,))
+        self.D = mx.ones(self.num_heads)
-        self.A_log = mx.log(mx.arange(1, self.num_heads + 1))
+        self.dt_bias = mx.zeros(self.num_heads)
-        self.D = mx.ones((self.num_heads,))
 
         self.norm = MambaRMSNormGated(self.intermediate_size, eps=args.layer_norm_epsilon)
 
         self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=args.use_bias)
 
-    def ssm_step(self, x, state=None):
+    def ssm_step(self, x, state, dt_proj):
         A = -mx.exp(self.A_log)
         D = self.D
-        deltaBC = self.x_proj(x)
+        delta = nn.softplus(dt_proj + self.dt_bias)
-        delta, B, C = mx.split(
+        
-            deltaBC,
+        B, C = mx.split(x, indices_or_sections=[self.state_size * self.n_groups], axis=-1)
-            indices_or_sections=[
+        
-                self.time_step_rank,
+        batch_size = B.shape[0]
-                self.time_step_rank + self.ssm_state_size,
+        B = B.reshape(batch_size, self.n_groups, self.state_size)
-            ],
+        C = C.reshape(batch_size, -1, self.state_size)
-            axis=-1,
+        
-        )
+        delta = delta.reshape(batch_size, self.num_heads, 1)
-        delta = nn.softplus(self.dt_proj(delta))
+        A = A.reshape(1, self.num_heads, 1)
-        new_state = mx.expand_dims(delta * x, -1) * mx.expand_dims(B, 1)
+        
-        if state is not None:
+        if state is None:
-            new_state += state * mx.exp(mx.expand_dims(delta, -1) * A)
+            new_state = delta * B
-        y = (new_state @ mx.expand_dims(C, -1)).squeeze(2)
+        else:
-        y = y + D * x
+            new_state = delta * (B + state * mx.exp(delta * A))
+        
+        y = mx.sum(new_state[:, :, None, :] * C[:, None, :, :], axis=(-1, -2))
+        y = y + D * x[:, :self.num_heads]
         return y, new_state
 
     def __call__(self, x, cache):
@@ -173,15 +181,28 @@ class Mamba2Mixer(nn.Module):
         for t in range(T):
             xt = x[:, t, :]
             xz = self.in_proj(xt)
-            x_t, z_t = xz.split(indices_or_sections=2, axis=1)
+            
+            x_t, z_t, dt_proj = mx.split(
+                xz,
+                indices_or_sections=[self.conv_dim, self.conv_dim + self.intermediate_size],
+                axis=-1
+            )
+
             conv_out, cache[0] = self.conv1d(mx.expand_dims(x_t, 1), cache[0])
             x_t = conv_out.squeeze(1)
             x_t = nn.silu(x_t)
-            y_t, cache[1] = self.ssm_step(x_t, cache[1])
+            y_t, cache[1] = self.ssm_step(x_t, cache[1], dt_proj)
             z_t = nn.silu(z_t)
-            output_t = y_t * z_t
+            
+            # Element-wise multiplication
+            output_t = y_t[:, :, None] * z_t[:, None, :]
+            
+            # Sum across the second dimension to match the intermediate_size
+            output_t = output_t.sum(axis=1)
+            
             output_t = self.out_proj(output_t)
             outputs.append(output_t)
+        
         output = mx.stack(outputs, axis=1)
         return output
 
@@ -240,6 +261,9 @@ class Model(nn.Module):
         else:
             logits = self.lm_head(x)
 
+        print(logits)
+        print(logits.shape)
+
         return logits
 
     def sanitize(self, weights):

llms/mlx_lm/models/mamba2.py

@@ -2,11 +2,13 @@
 
 import math
 from dataclasses import dataclass, field
-from typing import Tuple, Union
+from typing import Tuple, Union, Optional
 
-import mlx.core as mx
 import mlx.nn as nn
+import mlx.core as mx
+
 from .base import BaseModelArgs
+from .cache import Mamba2Cache
 
 # python -m mlx_lm.generate --model rokyang/mamba2-130m-hf  --prompt "hello how are you."
 
@@ -46,22 +48,6 @@ class ModelArgs(BaseModelArgs):
         if self.time_step_rank == "auto":
             self.time_step_rank = math.ceil(self.hidden_size / 16)
 
-
-class Mamba2Cache:
-    def __init__(self):
-        self.cache = [None, None]
-
-    def __setitem__(self, idx, value):
-        self.cache[idx] = value
-
-    def __getitem__(self, idx):
-        return self.cache[idx]
-
-    @property
-    def state(self):
-        return self.cache
-
-
 class MambaRMSNormGated(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
         super().__init__()
@@ -75,6 +61,7 @@ class MambaRMSNormGated(nn.Module):
         hidden_states = hidden_states * mx.rsqrt(variance + self.variance_epsilon)
         return self.weight * hidden_states
 
+
 class DepthWiseConv1d(nn.Module):
     def __init__(self, in_channels, out_channels, kernel_size, bias=True, groups=None, padding=0):
         super().__init__()
@@ -111,27 +98,22 @@ class DepthWiseConv1d(nn.Module):
 
 
 class Mamba2Mixer(nn.Module):
-    def __init__(self, args: ModelArgs):
+    def __init__(self, args, layer_idx):
         super().__init__()
-        self.args = args
+        self.layer_idx = layer_idx
-        self.intermediate_size = args.intermediate_size
-        self.time_step_rank = args.time_step_rank
-        self.conv_kernel_size = args.conv_kernel
         self.hidden_size = args.hidden_size
-        self.state_size = args.state_size
+        self.intermediate_size = args.intermediate_size
         self.num_heads = args.num_heads
-        self.head_dim = args.hidden_size // args.num_heads
+        self.head_dim = args.head_dim
+        self.ssm_state_size = args.state_size
         self.n_groups = args.n_groups
-
+        self.conv_kernel_size = args.conv_kernel
-        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.state_size
+        self.use_conv_bias = args.use_conv_bias
-        self.conv1d = DepthWiseConv1d(
+        self.use_bias = args.use_bias
-            in_channels=self.conv_dim,
+        self.time_step_min = args.time_step_min
-            out_channels=self.conv_dim,
+        self.time_step_max = args.time_step_max
-            bias=args.use_conv_bias,
+        self.chunk_size = args.chunk_size
-            kernel_size=args.conv_kernel,
+        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
-            groups=self.conv_dim,
-            padding=args.conv_kernel - 1
-        )
 
         projection_size = self.intermediate_size + self.conv_dim + self.num_heads
         self.in_proj = nn.Linear(
@@ -139,91 +121,151 @@ class Mamba2Mixer(nn.Module):
             projection_size,
             bias=args.use_bias
         )
-
+        self.conv1d = nn.Conv1d(
-        self.dt_bias = mx.ones((self.num_heads,))
+            self.conv_dim,
-        self.A_log = mx.log(mx.arange(1, self.num_heads + 1))
+            self.conv_dim,
-        self.D = mx.ones((self.num_heads,))
+            self.conv_kernel_size,
-
+            groups=self.conv_dim,
+            bias=self.use_conv_bias
+        )
+        self.act = nn.SiLU()
         self.norm = MambaRMSNormGated(self.intermediate_size, eps=args.layer_norm_epsilon)
+        self.out_proj = nn.Linear(
+            self.intermediate_size,
+            self.hidden_size,
+            bias=self.use_bias
+        )
 
-        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=args.use_bias)
+        self.A_log = mx.zeros(self.num_heads)
+        self.D = mx.ones(self.num_heads)
+        self.dt_bias = mx.zeros(self.num_heads)
+    
+    def __call__(self, input_states, cache):
+        batch_size, seq_len, _ = input_states.shape
+        dtype = input_states.dtype
 
-    def ssm_step(self, x, state, dt_proj):
+        projected_states = self.in_proj(input_states)
-        print(f"ssm_step input shapes - x: {x.shape}, dt_proj: {dt_proj.shape}")
-        A = -mx.exp(self.A_log)
-        D = self.D
-        delta = nn.softplus(dt_proj + self.dt_bias)
         
-        B, C = mx.split(x, indices_or_sections=[self.state_size * self.n_groups], axis=-1)
+        # Calculate the sizes of each split
-        print(f"ssm_step split shapes - B: {B.shape}, C: {C.shape}")
+        total_size = projected_states.shape[-1]
+        remaining_size = total_size - self.intermediate_size - self.conv_dim - self.num_heads
+        d_mlp = remaining_size // 2
+        sizes = [
+            d_mlp,
+            d_mlp,
+            self.intermediate_size,
+            self.conv_dim,
+            self.num_heads
+        ]
         
-        batch_size = B.shape[0]
+        # Perform the split operation
-        B = B.reshape(batch_size, self.n_groups, self.state_size)
+        split_result = mx.split(projected_states, sizes, axis=-1)
-        C = C.reshape(batch_size, -1, self.state_size)
-        print(f"After reshape - B: {B.shape}, C: {C.shape}")
         
-        delta = delta.reshape(batch_size, self.num_heads, 1)
+        # Print debug information
-        A = A.reshape(1, self.num_heads, 1)
+        print(f"Number of split parts: {len(split_result)}")
+        print(f"Shapes of split parts: {[part.shape for part in split_result]}")
         
-        if state is None:
+        # Flexibly handle the split result
-            new_state = delta * B
+        _, _, _, gate, hidden_states, dt = split_result
+
+        if cache is not None:
+            conv_state = cache.conv_states[self.layer_idx]
+            if conv_state is None:
+                # Initialize conv_state if it's None
+                conv_state = mx.zeros((batch_size, 1, self.conv_kernel_size, hidden_states.shape[-1]))
+            
+            conv_state = mx.roll(conv_state, -1, -2)  # Roll along the kernel dimension
+            
+            # Reshape hidden_states to match conv_state dimensions
+            hidden_states_reshaped = hidden_states[:, None, None, :]
+            
+            conv_state = mx.concat([conv_state[:, :, :-1, :], hidden_states_reshaped], axis=-2)
+            cache.conv_states[self.layer_idx] = conv_state
+            
+            # Adjust the convolution operation
+            hidden_states = mx.sum(conv_state * self.conv1d.weight[:, :, None, :], axis=(-2, -1))
+            
+            if self.use_conv_bias:
+                hidden_states += self.conv1d.bias
+            hidden_states = self.act(hidden_states)[:, None, :]
         else:
-            new_state = delta * (B + state * mx.exp(delta * A))
+            hidden_states = hidden_states.transpose(0, 2, 1)
-        
+            hidden_states = self.act(self.conv1d(hidden_states)).transpose(0, 2, 1)
-        print(f"Before final computation - new_state: {new_state.shape}, C: {C.shape}")
-        y = mx.sum(new_state[:, :, None, :] * C[:, None, :, :], axis=(-1, -2))
-        y = y + D * x[:, :self.num_heads]
-        print(f"ssm_step output shape - y: {y.shape}")
-        return y, new_state
 
-    def __call__(self, x, cache):
+        hidden_states, B, C = mx.split(hidden_states, [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size], axis=-1)
-        B, T, D = x.shape
-        print(f"__call__ input shape - x: {x.shape}")
-        if cache is None:
-            cache = [None, None]
 
-        outputs = []
+        A = -mx.exp(self.A_log.astype(mx.float32))
-        for t in range(T):
+        dt = nn.softplus(dt + self.dt_bias)
-            xt = x[:, t, :]
+        dt = mx.clip(dt, self.time_step_min, self.time_step_max)
-            xz = self.in_proj(xt)
-            print(f"After in_proj shape - xz: {xz.shape}")
-            
-            x_t, z_t, dt_proj = mx.split(
-                xz,
-                indices_or_sections=[self.conv_dim, self.conv_dim + self.intermediate_size],
-                axis=-1
-            )
-            print(f"After split shapes - x_t: {x_t.shape}, z_t: {z_t.shape}, dt_proj: {dt_proj.shape}")
 
-            conv_out, cache[0] = self.conv1d(mx.expand_dims(x_t, 1), cache[0])
+        hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).astype(mx.float32)
-            x_t = conv_out.squeeze(1)
+        B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).astype(mx.float32)
-            x_t = nn.silu(x_t)
+        C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).astype(mx.float32)
-            print(f"Before ssm_step shape - x_t: {x_t.shape}")
+
-            y_t, cache[1] = self.ssm_step(x_t, cache[1], dt_proj)
+        B = mx.repeat(B, repeats=self.num_heads // self.n_groups, axis=2)
-            z_t = nn.silu(z_t)
+        C = mx.repeat(C, repeats=self.num_heads // self.n_groups, axis=2)
-            print(f"After ssm_step shapes - y_t: {y_t.shape}, z_t: {z_t.shape}")
+
-            
+        if cache is not None and cache.seqlen_offset > 0:
-            # Element-wise multiplication
+            ssm_state = cache.ssm_states[self.layer_idx]
-            output_t = y_t[:, :, None] * z_t[:, None, :]
+            dA = mx.exp(dt[:, None, :, None] * A[None, :, None, None])
-            print(f"After multiplication shape - output_t: {output_t.shape}")
+            dB = dt[:, None, :, None] * B
-            
+            dBx = dB * hidden_states[:, :, :, None]
-            # Sum across the second dimension to match the intermediate_size
+            ssm_state = ssm_state * dA + dBx
-            output_t = output_t.sum(axis=1)
+            cache.ssm_states[self.layer_idx] = ssm_state
-            print(f"After sum shape - output_t: {output_t.shape}")
+
-            
+            y = mx.sum(ssm_state * C[:, None, :, :], axis=-1)
-            output_t = self.out_proj(output_t)
+            D = self.D[None, :, None].expand(self.D.shape[0], self.head_dim)
-            print(f"After out_proj shape - output_t: {output_t.shape}")
+            y = y + hidden_states * D
-            outputs.append(output_t)
+
-        
+            y = y.reshape(batch_size, -1)[:, None, :]
-        output = mx.stack(outputs, axis=1)
+        else:
-        print(f"Final output shape: {output.shape}")
+            # Implement chunked computation here (simplified version)
-        return output
+            pad_size = self.chunk_size - (seq_len % self.chunk_size)
+            hidden_states_padded = mx.pad(hidden_states, [(0, 0), (0, pad_size), (0, 0), (0, 0)])
+            B_padded = mx.pad(B, [(0, 0), (0, pad_size), (0, 0), (0, 0)])
+            C_padded = mx.pad(C, [(0, 0), (0, pad_size), (0, 0), (0, 0)])
+
+            chunks = seq_len // self.chunk_size + (1 if pad_size > 0 else 0)
+            y_list = []
+            ssm_state = mx.zeros((batch_size, self.num_heads, self.head_dim, self.ssm_state_size))
+
+            for i in range(chunks):
+                chunk_start = i * self.chunk_size
+                chunk_end = (i + 1) * self.chunk_size
+                chunk_h = hidden_states_padded[:, chunk_start:chunk_end]
+                chunk_B = B_padded[:, chunk_start:chunk_end]
+                chunk_C = C_padded[:, chunk_start:chunk_end]
+
+                chunk_dt = dt[:, chunk_start:chunk_end]
+                dA = mx.exp(chunk_dt[:, :, None, None] * A[None, None, :, None])
+                dB = chunk_dt[:, :, None, None] * chunk_B
+                dBx = dB * chunk_h[:, :, :, None]
+
+                chunk_y = mx.zeros_like(chunk_h)
+                for j in range(self.chunk_size):
+                    ssm_state = ssm_state * dA[:, j] + dBx[:, j]
+                    chunk_y[:, j] = mx.sum(ssm_state * chunk_C[:, j], axis=-1)
+
+                y_list.append(chunk_y)
+
+            y = mx.concat(y_list, axis=1)
+            if pad_size > 0:
+                y = y[:, :seq_len]
+
+            D = self.D[None, :, None].expand(self.D.shape[0], self.head_dim)
+            y = y + hidden_states * D
+            y = y.reshape(batch_size, seq_len, -1)
+
+        y = self.norm(y, gate)
+        contextualized_states = self.out_proj(y.astype(dtype))
+
+        return contextualized_states
 
 
 class Mamba2Block(nn.Module):
-    def __init__(self, args: ModelArgs):
+    def __init__(self, args: ModelArgs, layer_idx: int):
         super().__init__()
-        self.mixer = Mamba2Mixer(args)
+        self.mixer = Mamba2Mixer(args, layer_idx)
         self.norm = nn.RMSNorm(args.hidden_size)
 
     def __call__(self, x: mx.array, cache):
@@ -235,7 +277,7 @@ class Mamba2(nn.Module):
         super().__init__()
         self.args = args
         self.embeddings = nn.Embedding(args.vocab_size, args.hidden_size)
-        self.layers = [Mamba2Block(args) for idx in range(args.num_hidden_layers)]
+        self.layers = [Mamba2Block(args, idx) for idx in range(args.num_hidden_layers)]
         self.norm_f = nn.RMSNorm(args.hidden_size, eps=args.layer_norm_epsilon)
 
     def __call__(
@@ -274,6 +316,9 @@ class Model(nn.Module):
         else:
             logits = self.lm_head(x)
 
+        print(logits)
+        print(logits.shape)
+
         return logits
 
     def sanitize(self, weights):
@@ -282,8 +327,8 @@ class Model(nn.Module):
                 weights[k] = v.moveaxis(2, 1)
         return weights
 
-    def make_cache(self, batch_size: int = 1):
+    def make_cache(self):
-        return [Mamba2Cache() for _ in range(len(self.layers))]
+        return [Mamba2Cache(self.args.num_hidden_layers) for _ in range(len(self.layers))]
 
     @property
     def layers(self):