libmir-metal 0.3.0

Metal inference backend for libmir
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

use models::weights::{BitsAndBytesStorageDType, TensorBinding, TensorStorage};

use crate::engine::{Array, Dtype, Error, ModelTensors, Result, Stream};

#[derive(Debug)]
pub(in crate::engine) struct BitsAndBytes4BitLinear {
    weight: Array,
    absmax: Array,
    quant_map: Array,
    nested_absmax: Array,
    nested_quant_map: Array,
    input: usize,
    output: usize,
    block_size: usize,
    nested_block_size: Option<usize>,
    nested_offset_bits: u32,
}

pub(super) fn linear(
    tensors: &ModelTensors,
    binding: &TensorBinding,
) -> Result<BitsAndBytes4BitLinear> {
    let TensorStorage::BitsAndBytes4Bit {
        format,
        absmax,
        quant_map,
        nested_absmax,
        nested_quant_map,
        nested_offset_bits,
        ..
    } = &binding.storage
    else {
        return Err(invalid(binding, "requires bitsandbytes 4-bit storage"));
    };
    let Some([output, input]) = binding.logical_shape.as_deref() else {
        return Err(invalid(binding, "requires a logical matrix"));
    };
    let (output, input) = (*output, *input);
    let weight = tensors.get(&binding.source)?;
    require_bytes(&weight, output * input / 2, binding, "weight")?;
    require_dtype(&weight, storage_dtype(format.storage_dtype), binding, "weight")?;
    let absmax_array = tensors.get(absmax)?;
    let quant_map_array = tensors.get(quant_map)?;
    require(&quant_map_array, Dtype::Float32, &[16], binding, "quant map")?;
    let blocks = (output * input).div_ceil(format.block_size);
    let (nested_absmax_array, nested_quant_map_array) =
        if let Some(nested) = format.nested_block_size {
            require(&absmax_array, Dtype::Uint8, &[blocks], binding, "nested codes")?;
            let nested_absmax = tensors.get(
                nested_absmax
                    .as_deref()
                    .ok_or_else(|| invalid(binding, "nested absmax is missing"))?,
            )?;
            let nested_map = tensors.get(
                nested_quant_map
                    .as_deref()
                    .ok_or_else(|| invalid(binding, "nested map is missing"))?,
            )?;
            require(
                &nested_absmax,
                Dtype::Float32,
                &[blocks.div_ceil(nested)],
                binding,
                "nested absmax",
            )?;
            require(&nested_map, Dtype::Float32, &[256], binding, "nested map")?;
            (nested_absmax, nested_map)
        } else {
            require(&absmax_array, Dtype::Float32, &[blocks], binding, "absmax")?;
            (tensors.get(quant_map)?, tensors.get(quant_map)?)
        };
    Ok(BitsAndBytes4BitLinear {
        weight,
        absmax: absmax_array,
        quant_map: quant_map_array,
        nested_absmax: nested_absmax_array,
        nested_quant_map: nested_quant_map_array,
        input,
        output,
        block_size: format.block_size,
        nested_block_size: format.nested_block_size,
        nested_offset_bits: nested_offset_bits.unwrap_or(0),
    })
}

impl BitsAndBytes4BitLinear {
    pub(in crate::engine) fn forward(&self, input: &Array, stream: &Stream) -> Result<Array> {
        if !matches!(input.dtype()?, Dtype::Float16 | Dtype::Bfloat16) {
            return Err(Error::InvalidQuantization(
                "bitsandbytes input must be F16 or BF16".into(),
            ));
        }
        stream.kernels().bitsandbytes_4bit_linear(
            [
                input,
                &self.weight,
                &self.absmax,
                &self.quant_map,
                &self.nested_absmax,
                &self.nested_quant_map,
            ],
            self.input,
            self.output,
            self.block_size,
            self.nested_block_size,
            self.nested_offset_bits,
            stream,
        )
    }
}

fn storage_dtype(dtype: BitsAndBytesStorageDType) -> Dtype {
    match dtype {
        BitsAndBytesStorageDType::U8 => Dtype::Uint8,
        BitsAndBytesStorageDType::F16 => Dtype::Float16,
        BitsAndBytesStorageDType::Bf16 => Dtype::Bfloat16,
        BitsAndBytesStorageDType::F32 => Dtype::Float32,
    }
}

fn require(
    array: &Array,
    dtype: Dtype,
    shape: &[usize],
    binding: &TensorBinding,
    kind: &str,
) -> Result<()> {
    let expected = shape
        .iter()
        .copied()
        .map(i32::try_from)
        .collect::<std::result::Result<Vec<_>, _>>()?;
    if array.dtype()? == dtype && array.shape()? == expected {
        Ok(())
    } else {
        Err(invalid(binding, &format!("{kind} dtype or shape differs")))
    }
}

fn require_dtype(array: &Array, dtype: Dtype, binding: &TensorBinding, kind: &str) -> Result<()> {
    if array.dtype()? == dtype {
        Ok(())
    } else {
        Err(invalid(binding, &format!("{kind} dtype differs")))
    }
}

fn require_bytes(array: &Array, bytes: usize, binding: &TensorBinding, kind: &str) -> Result<()> {
    if array.byte_len()? == bytes {
        Ok(())
    } else {
        Err(invalid(binding, &format!("{kind} byte size differs")))
    }
}

fn invalid(binding: &TensorBinding, reason: &str) -> Error {
    Error::InvalidQuantization(format!("{}: {reason}", binding.source))
}