1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
//! GPU-accelerated quantized embedding table lookup.
//!
//! Supports 4-bit and 6-bit quantized embedding tables, performing
//! on-the-fly dequantization during gather. The dequantization formula
//! is `float_val = uint_val * scale + bias` with bf16 scales and biases.
use metal::MTLSize;
use crate::buffer::MlxBuffer;
use crate::encoder::CommandEncoder;
use crate::error::{MlxError, Result};
use crate::kernel_registry::KernelRegistry;
use super::encode_helpers::{as_bytes, encode_with_args, KernelArg};
/// Parameters for quantized embedding gather.
pub struct EmbeddingGatherParams {
/// Embedding dimension (number of float values per token).
pub embed_dim: usize,
/// Number of elements per quantization group (typically 64).
pub group_size: usize,
/// Quantization bit width: 4 or 6.
pub bits: u8,
/// Number of tokens to gather.
pub n_tokens: usize,
}
/// MSL-compatible parameter struct for the embedding kernel.
///
/// Must match the `EmbeddingParams` struct in `embedding.metal`.
#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct GpuEmbeddingParams {
embed_dim: u32,
group_size: u32,
packed_row_stride: u32,
n_groups_per_row: u32,
}
/// Encode a quantized embedding gather operation into the command buffer.
///
/// Looks up `n_tokens` rows from a quantized embedding table, dequantizing
/// each row on-the-fly on the GPU.
///
/// # Buffer expectations
///
/// * `weight_packed` — Packed quantized embedding table.
/// - 4-bit: `[vocab_size, embed_dim / 8]` uint32 values (8 values per uint32).
/// - 6-bit: `[vocab_size, embed_dim * 3 / 4]` uint8 bytes (4 values per 3 bytes).
/// * `scales` — bf16 scales, `[vocab_size, n_groups_per_row]`.
/// * `biases` — bf16 biases, `[vocab_size, n_groups_per_row]`.
/// * `token_ids` — uint32 token IDs, `[n_tokens]`.
/// * `output` — f32 output buffer, `[n_tokens, embed_dim]`.
///
/// # Errors
///
/// Returns `MlxError::InvalidArgument` if:
/// * `bits` is not 4 or 6
/// * `embed_dim` is zero
/// * `group_size` is zero
/// * `embed_dim` is not divisible by `group_size`
/// * `n_tokens` is zero
/// * Output buffer is too small
#[allow(clippy::too_many_arguments)]
pub fn embedding_gather(
encoder: &mut CommandEncoder,
registry: &mut KernelRegistry,
device: &metal::DeviceRef,
weight_packed: &MlxBuffer,
scales: &MlxBuffer,
biases: &MlxBuffer,
token_ids: &MlxBuffer,
output: &MlxBuffer,
params: &EmbeddingGatherParams,
) -> Result<()> {
// --- Validation ---
if params.bits != 4 && params.bits != 6 {
return Err(MlxError::InvalidArgument(format!(
"embedding_gather: bits must be 4 or 6, got {}",
params.bits
)));
}
if params.embed_dim == 0 {
return Err(MlxError::InvalidArgument(
"embedding_gather: embed_dim must be > 0".into(),
));
}
if params.group_size == 0 {
return Err(MlxError::InvalidArgument(
"embedding_gather: group_size must be > 0".into(),
));
}
if params.embed_dim % params.group_size != 0 {
return Err(MlxError::InvalidArgument(format!(
"embedding_gather: embed_dim ({}) must be divisible by group_size ({})",
params.embed_dim, params.group_size
)));
}
if params.n_tokens == 0 {
return Err(MlxError::InvalidArgument(
"embedding_gather: n_tokens must be > 0".into(),
));
}
let expected_output_bytes = params.n_tokens * params.embed_dim * std::mem::size_of::<f32>();
if output.byte_len() < expected_output_bytes {
return Err(MlxError::InvalidArgument(format!(
"embedding_gather: output buffer too small: need {} bytes, have {}",
expected_output_bytes,
output.byte_len()
)));
}
// --- Compute layout parameters ---
let n_groups_per_row = params.embed_dim / params.group_size;
let packed_row_stride: u32 = match params.bits {
4 => {
// 8 values per uint32; stride in uint32 count
(params.embed_dim / 8) as u32
}
6 => {
// 4 values per 3 bytes; stride in bytes
(params.embed_dim * 3 / 4) as u32
}
_ => unreachable!(), // validated above
};
let gpu_params = GpuEmbeddingParams {
embed_dim: params.embed_dim as u32,
group_size: params.group_size as u32,
packed_row_stride,
n_groups_per_row: n_groups_per_row as u32,
};
// --- Select kernel ---
let kernel_name = match params.bits {
4 => "embedding_gather_4bit",
6 => "embedding_gather_6bit",
_ => unreachable!(),
};
let pipeline = registry.get_pipeline(kernel_name, device)?;
// --- Encode dispatch ---
let grid = MTLSize::new(params.embed_dim as u64, params.n_tokens as u64, 1);
let tg_size = MTLSize::new(
std::cmp::min(256, params.embed_dim as u64),
1,
1,
);
let params_bytes = as_bytes(&gpu_params);
encode_with_args(
encoder,
pipeline,
&[
(0, KernelArg::Buffer(weight_packed)),
(1, KernelArg::Buffer(scales)),
(2, KernelArg::Buffer(biases)),
(3, KernelArg::Buffer(token_ids)),
(4, KernelArg::Buffer(output)),
(5, KernelArg::Bytes(params_bytes)),
],
grid,
tg_size,
);
Ok(())
}