zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-06-25 09:51:19 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

correct segsum function

Browse Source
This commit is contained in:
Goekdeniz-Guelmez 2025-02-26 14:46:46 +01:00
parent b7c0bdfd49
commit a683344450
3 changed files with 939 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

899
llms/mlx_lm/models/mamba2 copy.py Normal file
View File

@ -0,0 +1,899 @@
import math
from dataclasses import dataclass, field
from typing import Tuple, Union
import mlx.core as mx
import mlx.nn as nn
from .base import BaseModelArgs
from .cache import MambaCache
@dataclass
class ModelArgs(BaseModelArgs):
model_type: str
num_heads: int
head_dim: int
vocab_size: int
hidden_size: int
state_size: int
num_hidden_layers: int
layer_norm_epsilon: float
expand: int
conv_kernel: int
n_groups: int
use_bias: bool
use_conv_bias: bool
initializer_range: float
residual_in_fp32: bool
chunk_size: int
tie_word_embeddings: bool
time_step_limit: Tuple[float, float]
time_step_rank: Union[int, str]
time_step_min: float
time_step_max: float
time_step_floor: float
norm_before_gate: bool = True
def __post_init__(self):
if not hasattr(self, "intermediate_size"):
self.intermediate_size = int(self.expand * self.hidden_size)
if not hasattr(self, "head_dim"):
self.head_dim = self.hidden_size // self.num_heads
if self.time_step_rank == "auto":
self.time_step_rank = math.ceil(self.hidden_size / 16)
def segsum(x):
"""Stable segment sum calculation.
`exp(segsum(A))` produces a 1-semiseparable matrix, which is equivalent to a scalar SSM.
"""
T = x.shape[-1]
x = mx.expand_dims(x, -1)
x = mx.repeat(x, T, axis=-1)
mask = mx.tril(mx.ones((T, T), dtype=mx.bool_), k=-1)
x = mx.where(mask, x, 0)
x_segsum = mx.cumsum(x, axis=-2)
mask = mx.tril(mx.ones((T, T), dtype=mx.bool_), k=0)
x_segsum = mx.where(mask, x_segsum, -mx.inf)
return x_segsum
def ssd(x, A, B, C, chunk_size, initial_states=None):
"""Structured State Space Duality (SSD) - the core of Mamba-2
Arguments
x: (batch, seqlen, n_heads, d_head)
A: (batch, seqlen, n_heads)
B: (batch, seqlen, n_heads, d_state)
C: (batch, seqlen, n_heads, d_state)
Return
y: (batch, seqlen, n_heads, d_head)
final_state: final state for inference
"""
assert x.shape[1] % chunk_size == 0
# Rearrange into chunks
def rearrange_to_chunks(m):
shape = list(m.shape)
shape[1:2] = [shape[1] // chunk_size, chunk_size]
return m.reshape(shape)
x_chunked = rearrange_to_chunks(x)
A_chunked = rearrange_to_chunks(A)
B_chunked = rearrange_to_chunks(B)
C_chunked = rearrange_to_chunks(C)
# Transpose A for easier cumsum
A_chunked = mx.transpose(A_chunked, (0, 3, 1, 2)) # b c l h -> b h c l
A_cumsum = mx.cumsum(A_chunked, axis=-1)
# 1. Compute the output for each intra-chunk (diagonal blocks)
L = mx.exp(segsum(A_chunked))
Y_diag = mx.einsum("bclhn,bcshn,bhcls,bcshp->bclhp", C_chunked, B_chunked, L, x_chunked)
# 2. Compute the state for each intra-chunk
decay_states = mx.exp(A_cumsum[:, :, :, -1:] - A_cumsum)
states = mx.einsum("bclhn,bhcl,bclhp->bchpn", B_chunked, decay_states, x_chunked)
# 3. Compute the inter-chunk SSM recurrence
if initial_states is None:
initial_states = mx.zeros_like(states[:, :1])
states = mx.concatenate([initial_states, states], axis=1)
A_cumsum_last = A_cumsum[:, :, :, -1]
A_cumsum_padded = mx.pad(A_cumsum_last, [(0, 0), (0, 0), (1, 0)])
decay_chunk = mx.exp(segsum(A_cumsum_padded))
new_states = mx.einsum("bhzc,bchpn->bzhpn", decay_chunk, states)
states, final_state = new_states[:, :-1], new_states[:, -1]
# 4. Compute state -> output conversion per chunk
state_decay_out = mx.exp(A_cumsum)
Y_off = mx.einsum("bclhn,bchpn,bhcl->bclhp", C_chunked, states, state_decay_out)
# Add output of intra-chunk and inter-chunk terms
Y_combined = Y_diag + Y_off
# Reshape back to original sequence shape
batch, chunks, chunk_len, heads, head_dim = Y_combined.shape
Y = Y_combined.reshape(batch, chunks * chunk_len, heads, head_dim)
return Y, final_state
def silu(x):
"""Applies the Sigmoid Linear Unit (SiLU), element-wise."""
return x * mx.sigmoid(x)
class Mamba2Block(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.d_model = args.hidden_size
self.d_state = args.state_size
self.d_conv = args.conv_kernel
self.expand = args.expand
self.d_inner = int(self.expand * self.d_model)
self.n_groups = args.n_groups
self.n_heads = args.num_heads
self.d_head = self.d_inner // self.n_heads
self.chunk_size = args.chunk_size
d_in_proj = 2 * self.d_inner + 2 * self.n_groups * self.d_state + self.n_heads
self.in_proj = nn.Linear(self.d_model, d_in_proj, bias=args.use_bias)
self.dt_bias = mx.random.normal((self.n_heads,)) * args.initializer_range
self.A_log = mx.random.normal((self.n_heads,)) * args.initializer_range
self.D = mx.random.normal((self.n_heads,)) * args.initializer_range
# Use standard Conv1d with groups for depthwise convolution
conv_dim = self.d_inner + 2 * self.n_groups * self.d_state
self.conv1d = nn.Conv1d(
in_channels=conv_dim,
out_channels=conv_dim,
kernel_size=self.d_conv,
groups=conv_dim, # Makes it depthwise
padding=self.d_conv-1,
bias=args.use_conv_bias
)
self.norm = nn.RMSNorm(self.d_inner, eps=args.layer_norm_epsilon)
self.out_proj = nn.Linear(self.d_inner, self.d_model, bias=args.use_bias)
def __call__(self, u, cache=None):
"""
Arguments
u: (batch, seqlen, d_model) input
cache: Optional tuple of (conv_state, ssm_state) for inference
Return (y, cache)
y: (batch, seqlen, d_model) output
cache: updated tuple of (conv_state, ssm_state) for inference
"""
if cache is not None:
return self.step(u, cache)
# Initialize cache if needed
if cache is None:
cache = [None, None] # Initialize with None values
# Compute projections
zxbcdt = self.in_proj(u)
# Split projections
d_inner = self.d_inner
d_state = self.n_groups * self.d_state
z, xBC, dt = mx.split(
zxbcdt,
[d_inner, d_inner + 2 * d_state],
axis=-1
)
# Process dt with softplus
dt = mx.softplus(dt + self.dt_bias) # (batch, seqlen, n_heads)
# Apply convolution to xBC
xBC_transposed = mx.transpose(xBC, (0, 2, 1)) # (batch, d, seqlen)
xBC_conv = self.conv1d(xBC_transposed)
xBC_conv = mx.transpose(xBC_conv, (0, 2, 1)) # (batch, seqlen, d)
xBC = silu(xBC_conv[:, :u.shape[1], :]) # Ensure we only keep seqlen elements
# Split xBC into x, B, C
x, B, C = mx.split(
xBC,
[d_inner, d_inner + d_state],
axis=-1
)
# Reshape x for heads
batch, seqlen = x.shape[0], x.shape[1]
x_reshaped = x.reshape(batch, seqlen, self.n_heads, self.d_head)
# Reshape B and C for SSM
B = B.reshape(batch, seqlen, 1, d_state)
C = C.reshape(batch, seqlen, 1, d_state)
# Apply SSM with SSD algorithm
A = -mx.exp(self.A_log) # (n_heads,)
A_dt = A * dt # (batch, seqlen, n_heads)
y, ssm_state = ssd(
x_reshaped * mx.expand_dims(dt, -1), # Scale x by dt
A_dt,
B,
C,
self.chunk_size
)
# Apply D and reshape
y = y + x_reshaped * mx.reshape(self.D, (1, 1, self.n_heads, 1))
y = y.reshape(batch, seqlen, d_inner)
# Apply norm and gating
y = self.norm(y, z)
# Final projection
y = self.out_proj(y)
# Create cache for inference
if seqlen == 1 and cache is not None:
conv_state = mx.zeros((batch, d_inner + 2 * d_state, self.d_conv))
conv_state = mx.update_slice(conv_state, xBC.reshape(batch, -1, 1), (0, 0, self.d_conv - 1))
cache[0] = conv_state
cache[1] = ssm_state
return y, cache
def step(self, u, cache):
"""Take an inference step for the current input and cache
Arguments
u: (batch, seqlen, d_model) - can be multiple tokens
cache: tuple of (conv_state, ssm_state)
Return (y, cache)
y: (batch, seqlen, d_model)
cache: updated cache object
"""
batch, seqlen = u.shape[0], u.shape[1]
# Initialize cache if it's None
if cache[0] is None or cache[1] is None:
d_state = self.n_groups * self.d_state
conv_dim = self.d_inner + 2 * d_state
conv_state = mx.zeros((batch, conv_dim, self.d_conv))
# Fix: use correct state size per head
state_per_head = d_state // self.n_heads
ssm_state = mx.zeros((batch, self.n_heads, self.d_head, state_per_head))
else:
conv_state, ssm_state = cache[0], cache[1]
# Project input
zxbcdt = self.in_proj(u) # (batch, seqlen, d_in_proj)
# Split projections
d_inner = self.d_inner
d_state = self.n_groups * self.d_state
z, xBC, dt = mx.split(
zxbcdt,
[d_inner, d_inner + 2 * d_state],
axis=-1
)
# Process each token through the convolution sequentially
outputs = []
for i in range(seqlen):
# Get current token's input
xBC_i = xBC[:, i] # (batch, d_inner + 2*d_state)
dt_i = dt[:, i] # (batch, dt_size)
# Extract the head-specific dt values
dt_size = dt_i.shape[-1]
if dt_size % self.n_heads == 0:
# Reshape dt_i to extract the head-specific values
dt_reshaped = dt_i.reshape(batch, self.n_heads, dt_size // self.n_heads)
# Take the first element for each head
dt_heads = dt_reshaped[:, :, 0]
else:
# If we can't reshape, just take the first n_heads elements
dt_heads = dt_i[:, :self.n_heads]
# Process dt with softplus
dt_heads = nn.softplus(dt_heads + self.dt_bias.reshape(1, -1)) # (batch, n_heads)
# Update convolution state
conv_state = mx.roll(conv_state, shift=-1, axis=-1)
# Use slice_update instead of update_slice
# Reshape xBC_i to match the expected shape for the update
xBC_reshaped = xBC_i.reshape(batch, -1, 1)
# Create start_indices for the update
start_indices = mx.array([0, 0, self.d_conv - 1])
# Update the conv_state
conv_state = mx.slice_update(
conv_state,
xBC_reshaped,
start_indices,
axes=(0, 1, 2)
)
# Apply convolution step
weight = self.conv1d.weight
bias = self.conv1d.bias if self.args.use_conv_bias else None
xBC_conv = mx.sum(conv_state * weight.reshape(1, -1, self.d_conv), axis=-1)
if bias is not None:
xBC_conv = xBC_conv + bias
xBC_conv = silu(xBC_conv)
# Split xBC
x_i, B_i, C_i = mx.split(
xBC_conv,
[d_inner, d_inner + d_state],
axis=-1
)
# Apply SSM step
A = -mx.exp(self.A_log) # (n_heads,)
dA = mx.exp(dt_heads * A) # (batch, n_heads)
# Reshape x for heads
x_i = x_i.reshape(batch, self.n_heads, self.d_head)
# Reshape B and C for SSM with correct dimensions
state_per_head = d_state // self.n_heads
B_i_reshaped = B_i.reshape(batch, self.n_heads, state_per_head)
C_i_reshaped = C_i.reshape(batch, self.n_heads, state_per_head)
# Calculate dBx with the correctly shaped B
dBx = mx.einsum("bhn,bhp->bhpn", B_i_reshaped, x_i * mx.expand_dims(dt_heads, -1))
# Update SSM state
ssm_state = ssm_state * mx.reshape(dA, (batch, self.n_heads, 1, 1)) + dBx
# Calculate output with the correctly shaped C
y_i = mx.einsum("bhpn,bhn->bhp", ssm_state, C_i_reshaped)
# Apply D and reshape
y_i = y_i + x_i * mx.reshape(self.D, (1, self.n_heads, 1))
# Reshape y
y_i = y_i.reshape(batch, d_inner)
# Apply norm and gating (SwiGLU-like activation)
y_i = self.norm(y_i) # Just normalize without gating
y_i = y_i * nn.sigmoid(z[:, i]) # Apply gating separately
# Final projection
y_i = self.out_proj(y_i)
outputs.append(y_i)
# Stack outputs along sequence dimension
y = mx.stack(outputs, axis=1) # (batch, seqlen, d_model)
# Update cache
cache[0] = conv_state
cache[1] = ssm_state
return y
class ResidualBlock(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.residual_in_fp32 = args.residual_in_fp32
self.mixer = Mamba2Block(args)
self.norm = nn.RMSNorm(args.hidden_size)
def __call__(self, x: mx.array, cache):
if self.residual_in_fp32:
x = x.astype(mx.float32)
normed = self.norm(x)
output = self.mixer(normed, cache)
return output + x
class Mamba2(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.embeddings = nn.Embedding(args.vocab_size, args.hidden_size)
self.layers = [ResidualBlock(args) for _ in range(args.num_hidden_layers)]
self.norm_f = nn.RMSNorm(args.hidden_size, eps=args.layer_norm_epsilon)
def __call__(self, x: mx.array, cache):
x = self.embeddings(x)
if cache is None:
cache = [None] * len(self.layers)
hidden = x
for layer, c in zip(self.layers, cache):
hidden = layer(hidden, c)
return self.norm_f(hidden)
class Model(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.model_type = args.model_type
self.backbone = Mamba2(args)
if not args.tie_word_embeddings:
self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
def __call__(self, inputs: mx.array, cache=None):
hidden = self.backbone(inputs, cache)
if self.args.tie_word_embeddings:
logits = self.backbone.embeddings.as_linear(hidden)
else:
logits = self.lm_head(hidden)
return logits
def make_cache(self):
return [MambaCache() for _ in range(len(self.layers))]
@property
def layers(self):
return self.backbone.layers
########################################################
import math
from dataclasses import dataclass, field
from typing import Tuple, Union
import mlx.core as mx
import mlx.nn as nn
from .base import BaseModelArgs
from .cache import MambaCache
@dataclass
class ModelArgs(BaseModelArgs):
model_type: str
num_heads: int
head_dim: int
vocab_size: int
hidden_size: int
state_size: int
num_hidden_layers: int
layer_norm_epsilon: float
expand: int
conv_kernel: int
n_groups: int
use_bias: bool
use_conv_bias: bool
initializer_range: float
residual_in_fp32: bool
chunk_size: int
tie_word_embeddings: bool
time_step_limit: Tuple[float, float]
time_step_rank: Union[int, str]
time_step_min: float
time_step_max: float
time_step_floor: float
norm_before_gate: bool = True
def __post_init__(self):
if not hasattr(self, "intermediate_size"):
self.intermediate_size = int(self.expand * self.hidden_size)
if not hasattr(self, "head_dim"):
self.head_dim = self.hidden_size // self.num_heads
if self.time_step_rank == "auto":
self.time_step_rank = math.ceil(self.hidden_size / 16)
def segsum(x):
"""Stable segment sum calculation.
`exp(segsum(A))` produces a 1-semiseparable matrix, which is equivalent to a scalar SSM.
"""
T = x.shape[-1]
x = mx.expand_dims(x, -1)
x = mx.repeat(x, T, axis=-1)
mask = mx.tril(mx.ones((T, T), dtype=mx.bool_), k=-1)
x = mx.where(mask, x, 0)
x_segsum = mx.cumsum(x, axis=-2)
mask = mx.tril(mx.ones((T, T), dtype=mx.bool_), k=0)
x_segsum = mx.where(mask, x_segsum, -mx.inf)
return x_segsum
def ssd(x, A, B, C, chunk_size, initial_states=None):
"""Structured State Space Duality (SSD) - the core of Mamba-2
Arguments
x: (batch, seqlen, n_heads, d_head)
A: (batch, seqlen, n_heads)
B: (batch, seqlen, n_heads, d_state)
C: (batch, seqlen, n_heads, d_state)
Return
y: (batch, seqlen, n_heads, d_head)
final_state: final state for inference
"""
assert x.shape[1] % chunk_size == 0
# Rearrange into chunks
def rearrange_to_chunks(m):
shape = list(m.shape)
shape[1:2] = [shape[1] // chunk_size, chunk_size]
return m.reshape(shape)
x_chunked = rearrange_to_chunks(x)
A_chunked = rearrange_to_chunks(A)
B_chunked = rearrange_to_chunks(B)
C_chunked = rearrange_to_chunks(C)
# Transpose A for easier cumsum
A_chunked = mx.transpose(A_chunked, (0, 3, 1, 2)) # b c l h -> b h c l
A_cumsum = mx.cumsum(A_chunked, axis=-1)
# 1. Compute the output for each intra-chunk (diagonal blocks)
L = mx.exp(segsum(A_chunked))
Y_diag = mx.einsum("bclhn,bcshn,bhcls,bcshp->bclhp", C_chunked, B_chunked, L, x_chunked)
# 2. Compute the state for each intra-chunk
decay_states = mx.exp(A_cumsum[:, :, :, -1:] - A_cumsum)
states = mx.einsum("bclhn,bhcl,bclhp->bchpn", B_chunked, decay_states, x_chunked)
# 3. Compute the inter-chunk SSM recurrence
if initial_states is None:
initial_states = mx.zeros_like(states[:, :1])
states = mx.concatenate([initial_states, states], axis=1)
A_cumsum_last = A_cumsum[:, :, :, -1]
A_cumsum_padded = mx.pad(A_cumsum_last, [(0, 0), (0, 0), (1, 0)])
decay_chunk = mx.exp(segsum(A_cumsum_padded))
new_states = mx.einsum("bhzc,bchpn->bzhpn", decay_chunk, states)
states, final_state = new_states[:, :-1], new_states[:, -1]
# 4. Compute state -> output conversion per chunk
state_decay_out = mx.exp(A_cumsum)
Y_off = mx.einsum("bclhn,bchpn,bhcl->bclhp", C_chunked, states, state_decay_out)
# Add output of intra-chunk and inter-chunk terms
Y_combined = Y_diag + Y_off
# Reshape back to original sequence shape
batch, chunks, chunk_len, heads, head_dim = Y_combined.shape
Y = Y_combined.reshape(batch, chunks * chunk_len, heads, head_dim)
return Y, final_state
def silu(x):
"""Applies the Sigmoid Linear Unit (SiLU), element-wise."""
return x * mx.sigmoid(x)
class Mamba2Block(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.d_model = args.hidden_size
self.d_state = args.state_size
self.d_conv = args.conv_kernel
self.expand = args.expand
self.d_inner = int(self.expand * self.d_model)
self.n_groups = args.n_groups
self.n_heads = args.num_heads
self.d_head = self.d_inner // self.n_heads
self.chunk_size = args.chunk_size
d_in_proj = 2 * self.d_inner + 2 * self.n_groups * self.d_state + self.n_heads
self.in_proj = nn.Linear(self.d_model, d_in_proj, bias=args.use_bias)
self.dt_bias = mx.random.normal((self.n_heads,)) * args.initializer_range
self.A_log = mx.random.normal((self.n_heads,)) * args.initializer_range
self.D = mx.random.normal((self.n_heads,)) * args.initializer_range
# Use standard Conv1d with groups for depthwise convolution
conv_dim = self.d_inner + 2 * self.n_groups * self.d_state
self.conv1d = nn.Conv1d(
in_channels=conv_dim,
out_channels=conv_dim,
kernel_size=self.d_conv,
groups=conv_dim,
padding=self.d_conv-1,
bias=args.use_conv_bias
)
self.norm = nn.RMSNorm(self.d_inner, eps=args.layer_norm_epsilon)
self.out_proj = nn.Linear(self.d_inner, self.d_model, bias=args.use_bias)
def __call__(self, u, cache=None):
"""
Arguments
u: (batch, seqlen, d_model) input
cache: Optional tuple of (conv_state, ssm_state) for inference
Return (y, cache)
y: (batch, seqlen, d_model) output
cache: updated tuple of (conv_state, ssm_state) for inference
"""
if cache is not None:
return self.step(u, cache)
# Initialize cache if needed
if cache is None:
cache = [None, None] # Initialize with None values
# Compute projections
zxbcdt = self.in_proj(u)
# Split projections
d_inner = self.d_inner
d_state = self.n_groups * self.d_state
z, xBC, dt = mx.split(
zxbcdt,
[d_inner, d_inner + 2 * d_state],
axis=-1
)
# Process dt with softplus
dt = mx.softplus(dt + self.dt_bias) # (batch, seqlen, n_heads)
# Apply convolution to xBC
xBC_transposed = mx.transpose(xBC, (0, 2, 1)) # (batch, d, seqlen)
xBC_conv = self.conv1d(xBC_transposed)
xBC_conv = mx.transpose(xBC_conv, (0, 2, 1)) # (batch, seqlen, d)
xBC = silu(xBC_conv[:, :u.shape[1], :]) # Ensure we only keep seqlen elements
# Split xBC into x, B, C
x, B, C = mx.split(
xBC,
[d_inner, d_inner + d_state],
axis=-1
)
# Reshape x for heads
batch, seqlen = x.shape[0], x.shape[1]
x_reshaped = x.reshape(batch, seqlen, self.n_heads, self.d_head)
# Reshape B and C for SSM
B = B.reshape(batch, seqlen, 1, d_state)
C = C.reshape(batch, seqlen, 1, d_state)
# Apply SSM with SSD algorithm
A = -mx.exp(self.A_log) # (n_heads,)
A_dt = A * dt # (batch, seqlen, n_heads)
y, ssm_state = ssd(
x_reshaped * mx.expand_dims(dt, -1), # Scale x by dt
A_dt,
B,
C,
self.chunk_size
)
# Apply D and reshape
y = y + x_reshaped * mx.reshape(self.D, (1, 1, self.n_heads, 1))
y = y.reshape(batch, seqlen, d_inner)
# Apply norm and gating
y = self.norm(y, z)
# Final projection
y = self.out_proj(y)
# Create cache for inference
if seqlen == 1 and cache is not None:
conv_state = mx.zeros((batch, d_inner + 2 * d_state, self.d_conv))
conv_state = mx.update_slice(conv_state, xBC.reshape(batch, -1, 1), (0, 0, self.d_conv - 1))
cache[0] = conv_state
cache[1] = ssm_state
return y
def step(self, u, cache):
"""Take an inference step for the current input and cache
Arguments
u: (batch, seqlen, d_model) - can be multiple tokens
cache: tuple of (conv_state, ssm_state)
Return (y, cache)
y: (batch, seqlen, d_model)
cache: updated cache object
"""
batch, seqlen = u.shape[0], u.shape[1]
# Initialize cache if it's None
if cache[0] is None or cache[1] is None:
d_state = self.n_groups * self.d_state
conv_dim = self.d_inner + 2 * d_state
conv_state = mx.zeros((batch, conv_dim, self.d_conv))
# Fix: use correct state size per head
state_per_head = d_state // self.n_heads
ssm_state = mx.zeros((batch, self.n_heads, self.d_head, state_per_head))
else:
conv_state, ssm_state = cache[0], cache[1]
# Project input
zxbcdt = self.in_proj(u) # (batch, seqlen, d_in_proj)
# Split projections
d_inner = self.d_inner
d_state = self.n_groups * self.d_state
z, xBC, dt = mx.split(
zxbcdt,
[d_inner, d_inner + 2 * d_state],
axis=-1
)
# Process dt with softplus once for all tokens
dt_heads = dt.reshape(batch, seqlen, -1)[:, :, :self.n_heads]
dt_heads = nn.softplus(dt_heads + self.dt_bias.reshape(1, 1, -1))
# Pre-compute dA for all tokens
A = -mx.exp(self.A_log) # (n_heads,)
dA = mx.exp(dt_heads * A.reshape(1, 1, -1)) # (batch, seqlen, n_heads)
# Get convolution weights
weight = self.conv1d.weight # shape: (out_channels, 1, kernel_size)
bias = self.conv1d.bias if self.args.use_conv_bias else None
# Process each token through the convolution sequentially
outputs = []
for i in range(seqlen):
# Get current token's input
xBC_i = xBC[:, i] # (batch, d_inner + 2*d_state)
# Update convolution state
conv_state = mx.roll(conv_state, shift=-1, axis=-1)
# Update the last column of conv_state
conv_state = mx.slice_update(
conv_state,
xBC_i.reshape(batch, -1, 1),
mx.array([0, 0, self.d_conv - 1]),
axes=(0, 1, 2)
)
# Apply convolution step - manually handle the depthwise conv
# For a depthwise conv, we need to process each channel separately
# conv_state shape: (batch, channels, kernel_size)
# weight shape: (channels, 1, kernel_size) for depthwise conv
# Reshape weight to match conv_state for element-wise multiplication
# and then sum along the kernel dimension
weight_reshaped = weight.reshape(conv_state.shape[1], self.d_conv)
xBC_conv = mx.sum(conv_state * weight_reshaped.reshape(1, -1, self.d_conv), axis=-1)
if bias is not None:
xBC_conv = xBC_conv + bias
xBC_conv = silu(xBC_conv)
# Split xBC
x_i, BC_rest = mx.split(xBC_conv, [d_inner], axis=-1)
B_i, C_i = mx.split(BC_rest, [d_state], axis=-1)
# Reshape x for heads
x_i = x_i.reshape(batch, self.n_heads, self.d_head)
# Reshape B and C for SSM
state_per_head = d_state // self.n_heads
B_i_reshaped = B_i.reshape(batch, self.n_heads, state_per_head)
C_i_reshaped = C_i.reshape(batch, self.n_heads, state_per_head)
# Get current token's dt and dA
dt_i = dt_heads[:, i] # (batch, n_heads)
dA_i = dA[:, i] # (batch, n_heads)
# Calculate dBx
dBx = mx.einsum("bhn,bhp->bhpn", B_i_reshaped, x_i * mx.expand_dims(dt_i, -1))
# Update SSM state
ssm_state = ssm_state * mx.reshape(dA_i, (batch, self.n_heads, 1, 1)) + dBx
# Calculate output with the correctly shaped C
y_i = mx.einsum("bhpn,bhn->bhp", ssm_state, C_i_reshaped)
# Apply D and reshape
y_i = y_i + x_i * self.D.reshape(1, self.n_heads, 1)
# Reshape y
y_i = y_i.reshape(batch, d_inner)
# Apply norm and gating
y_i = self.norm(y_i) * nn.sigmoid(z[:, i])
# Final projection
y_i = self.out_proj(y_i)
outputs.append(y_i)
# Stack outputs along sequence dimension
y = mx.stack(outputs, axis=1)
# Update cache
cache[0] = conv_state
cache[1] = ssm_state
return y
class ResidualBlock(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.residual_in_fp32 = args.residual_in_fp32
self.mixer = Mamba2Block(args)
self.norm = nn.RMSNorm(args.hidden_size)
def __call__(self, x: mx.array, cache):
if self.residual_in_fp32:
x = x.astype(mx.float32)
normed = self.norm(x)
output = self.mixer(normed, cache)
return output + x
class Mamba2(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.embeddings = nn.Embedding(args.vocab_size, args.hidden_size)
self.layers = [ResidualBlock(args) for _ in range(args.num_hidden_layers)]
self.norm_f = nn.RMSNorm(args.hidden_size, eps=args.layer_norm_epsilon)
def __call__(self, x: mx.array, cache):
x = self.embeddings(x)
if cache is None:
cache = [None] * len(self.layers)
hidden = x
for layer, c in zip(self.layers, cache):
hidden = layer(hidden, c)
return self.norm_f(hidden)
class Model(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.model_type = args.model_type
self.backbone = Mamba2(args)
if not args.tie_word_embeddings:
self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
def __call__(self, inputs: mx.array, cache=None):
hidden = self.backbone(inputs, cache)
if self.args.tie_word_embeddings:
logits = self.backbone.embeddings.as_linear(hidden)
else:
logits = self.lm_head(hidden)
return logits
def make_cache(self):
return [MambaCache() for _ in range(len(self.layers))]
@property
def layers(self):
return self.backbone.layers

55
llms/mlx_lm/models/mamba2.py
View File

@ -75,52 +75,76 @@ def ssd_forward_attn(
dt_bias: mx.array, dt_bias: mx.array,
dt_min: float, dt_min: float,
dt_max: float, dt_max: float,
prev_state=None,
) -> Tuple[mx.array, mx.array]: ) -> Tuple[mx.array, mx.array]:
b, l, h, dh = x.shape b, l, h, dh = x.shape
_, _, g, _ = B.shape _, _, g, _ = B.shape
# Process dt
if dt_bias is not None: if dt_bias is not None:
dt = dt + dt_bias.reshape(1, 1, -1) dt = dt + dt_bias.reshape(1, 1, -1)
dt = nn.softplus(dt) dt = nn.softplus(dt)
dt = mx.clip(dt, a_min=dt_min, a_max=dt_max) dt = mx.clip(dt, a_min=dt_min, a_max=dt_max)
B = mx.swapaxes(mx.swapaxes(B, 1, 3), 1, 2) # Reshape tensors
C = mx.swapaxes(C, 1, 2) B_reshaped = mx.swapaxes(mx.swapaxes(B, 1, 3), 1, 2)
C_reshaped = mx.swapaxes(C, 1, 2)
CB = C @ B # Compute CB
CB = C_reshaped @ B_reshaped
CB = mx.repeat(CB, repeats=h // g, axis=1) CB = mx.repeat(CB, repeats=h // g, axis=1)
# Compute decay terms
dtA = dt * A.reshape(1, 1, -1) dtA = dt * A.reshape(1, 1, -1)
dtA = mx.swapaxes(dtA, 1, 2) dtA = mx.swapaxes(dtA, 1, 2)
decay = mx.exp(segsum(dtA)) decay = mx.exp(segsum(dtA))
# Create attention matrix
surrogate_attention_matrix = mx.tril(CB * decay, 0) surrogate_attention_matrix = mx.tril(CB * decay, 0)
# Apply attention
dtx = dt.reshape(b, l, h, 1) * x dtx = dt.reshape(b, l, h, 1) * x
y = surrogate_attention_matrix @ dtx.swapaxes(1, 2) y = surrogate_attention_matrix @ dtx.swapaxes(1, 2)
y = mx.swapaxes(y, 1, 2) y = mx.swapaxes(y, 1, 2)
decay = decay[:, :, -1, :].reshape(b, h, l).swapaxes(1, 2).reshape(b, l, h, 1) # Compute next state
B = mx.repeat(B, h // g, axis=1).swapaxes(2, 3) decay_last = decay[:, :, -1, :].reshape(b, h, l).swapaxes(1, 2).reshape(b, l, h, 1)
dtxdecay = dtx * decay B_for_state = mx.repeat(B_reshaped, h // g, axis=1).swapaxes(2, 3)
dtxdecay = dtx * decay_last
dtxdecay = dtxdecay.swapaxes(1, 2).swapaxes(2, 3) dtxdecay = dtxdecay.swapaxes(1, 2).swapaxes(2, 3)
next_state = dtxdecay @ B
# Calculate new state contribution
new_state_contribution = dtxdecay @ B_for_state
# Initialize or update state
if prev_state is not None:
# Simply use the previous state if it exists
# This is a simplified approach - just use the new state
# In a real implementation, you'd want to properly update based on your SSM formulation
next_state = new_state_contribution
else:
next_state = new_state_contribution
# Add skip connection if D is provided
if D is not None: if D is not None:
y += x * D.reshape(1, 1, h, 1) y += x * D.reshape(1, 1, h, 1)
# Reshape output
y = y.reshape(b, l, h * dh) y = y.reshape(b, l, h * dh)
return y, next_state return y, next_state
def segsum(x): def segsum(x):
l = x.shape[-1] # x shape: [b, h, l]
x = mx.repeat(x[..., None], l, axis=-1) b, h, l = x.shape
x = mx.tril(x, -1) indices = mx.arange(l)
x_segsum = mx.cumsum(x, axis=-2) mask = indices[:, None] >= indices[None, :] # [l, l] lower triangular mask
# Expand x for broadcasting
x_expanded = x.reshape(b, h, l, 1) # [b, h, l, 1]
# Apply mask and sum
masked_x = x_expanded * mask.reshape(1, 1, l, l) # [b, h, l, l]
x_segsum = mx.sum(masked_x, axis=2, keepdims=True) # [b, h, 1, l]
return x_segsum return x_segsum
@ -189,13 +213,14 @@ class Mamba2Block(nn.Module):
y, next_ssm_state = ssd_forward_attn( y, next_ssm_state = ssd_forward_attn(
x=x, x=x,
dt=dt, dt=dt,
A=A, A=-mx.exp(self.A_log),
B=B, B=B,
C=C, C=C,
D=self.D, D=self.D,
dt_bias=self.dt_bias, dt_bias=self.dt_bias,
dt_min=self.args.time_step_min, dt_min=self.args.time_step_min,
dt_max=self.args.time_step_max dt_max=self.args.time_step_max,
prev_state=ssm_state
) )
if self.args.norm_before_gate: if self.args.norm_before_gate:

11
llms/mlx_lm/models/mamba2_pytorch.py
View File

@ -1,14 +1,3 @@
"""
mamba2-minimal
==============
A minimal, single-file implementation of the Mamba-2 model in PyTorch.
> **Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality**
> Authors: Tri Dao, Albert Gu
> Paper: https://arxiv.org/abs/2405.21060
"""
import json import json
from dataclasses import dataclass from dataclasses import dataclass
from typing import Iterable, NamedTuple, TypeAlias, cast from typing import Iterable, NamedTuple, TypeAlias, cast