Skip to content

Weight-Quantized Matmul

mm_weight_quant and addmm_weight_quant consume a host-packed weight blob. hart.weight_converter prepares the required opaque buffer from the weights, quantization parameters and optional bias. Legato streams the opaque rank-1 i8 payload to mpu_packed_buffer and performs the matrix multiplication on the EVT0 MPU.

Verification scope

The complete example below checks converter byte counts and compiles four variants through BINARY with launch=False. It does not claim FP8 numerical parity on FSim or hardware. The current EVT0 Torch FSim quantization predicate omits FP8/FP4, so packed scale metadata is not processed correctly. An FP8 input-channel case can fail with incompatible scale shapes; even an output-channel launch without an exception is not a numerical correctness test. See Lagora FP8 grouping issue.

This is a simulator limitation; the documentation does not implement its fix. Do not turn on launch for these samples as an FSim validation shortcut.

Granularity and shape

granularity Scale shape for [C, M] weights quant_group_size
"per_output_channel" [M] ignored
"per_input_channel" [C / group_size, M] actual C-group size, default 128

For input-channel grouping, the group size must be at least 128, a multiple of 128, and divide C. The group-size argument must match the converter configuration. Bias has shape [M]; its pre-packing dtype is passed separately to addmm_weight_quant.

Packing and compiling

Download weight_quant_matmul.py, or run the repository files with the native compiler/toolchain and hart.weight_converter available:

1
HA_TOOLCHAIN=1 python "$DOCS_ROOT/docs/legato-language/tutorials/code/weight_quant_matmul.py"

weight_i8 stores FP8 E4M3FN bit patterns produced by .to(torch.float8_e4m3fn).view(torch.int8). Random signed integers are not an FP8 quantization procedure and may encode NaNs. scale_f32 remains FP32. The converter's bytearray buffers stay alive until convert() returns; passing temporary buffers through raw pointers can otherwise dangle.

Every kernel has an output parameter and explicitly stores the result. CPU arguments are used only to compile these typed kernels, never for a device launch. The four byte counts (8704, 8704, 16896, 9216) are checked against actual converter output for the shown configurations; do not reuse them for other shapes or quantization formats.

  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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
"""Pack valid FP8 payloads and compile all quantized matmul variants (no FSim launch)."""

import os
from pathlib import Path

import torch
from hart.weight_converter import (
    ArchitectureType,
    DataType,
    QuantType,
    Shape,
    Tensor,
    TensorType,
    WeightConverterConfig,
    convert,
)

import legato
import legato.model.bertha as bertha

# Store compiler artifacts and optional simulator logs in a configurable directory.
OUTPUT = Path(os.environ.get("LEGATO_EXAMPLE_OUTPUT", "legato_example_output"))


def _to_bytes(tensor: torch.Tensor) -> bytearray:
    return bytearray(tensor.contiguous().cpu().numpy().tobytes())


def pack_fp8_weight(
    weight_i8: torch.Tensor,
    scale_f32: torch.Tensor,
    *,
    quant_type: QuantType,
    group_size: int = 128,
    bias_f32: torch.Tensor | None = None,
) -> torch.Tensor:
    """Pack an FP8 EVT0 weight.

    `weight_i8` is `[C, M]`; `scale_f32` is `[M]` for output-channel-wise
    or `[C / group_size, M]` for input-channel-wise.
    """
    input_channels, output_channels = weight_i8.shape
    cfg = WeightConverterConfig(
        ArchitectureType.EVT0,
        quant_type,
        group_size if quant_type == QuantType.InputChannelWiseQuant else 0,
    )
    num_groups = input_channels // group_size if quant_type == QuantType.InputChannelWiseQuant else 1
    weight_buf = _to_bytes(weight_i8)
    scale_buf = _to_bytes(scale_f32)
    weight = Tensor(
        weight_buf,
        Shape(1, 1, input_channels, output_channels),
        DataType.FP8,
        TensorType.Weight,
    )
    scale = Tensor(
        scale_buf,
        Shape(1, 1, num_groups, output_channels),
        DataType.FP32,
        TensorType.Scale,
    )
    bias = None
    bias_buf = None
    if bias_f32 is not None:
        bias_buf = _to_bytes(bias_f32)
        bias = Tensor(
            bias_buf,
            Shape(1, 1, 1, output_channels),
            DataType.FP32,
            TensorType.Bias,
        )
    packed = convert(cfg, weight, bias, scale, None, None, None, output_data_type=DataType.BF16)
    return torch.frombuffer(bytearray(packed), dtype=torch.int8).clone()


@legato.compile(
    backend=lambda ctx: bertha.Bertha(ctx, "bertha"), param_kinds={"a": "universal", "w": "top", "out": "core"}
)
def mm_weight_quant_oc(
    a: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 128], "shared_memory"),
    w: legato.get_model().tensor_type(legato.types.integer("int8"), [8704], "shared_dram"),
    out: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 64], "shared_dram"),
):
    """Compile an output-channel-wise FP8 packed-weight matmul."""
    device = legato.get_context().get_device()
    with device.get_top():
        legato.tensor.request_load(a, device.get_core(0), "sram")
        legato.tensor.request_load(w, device.get_core(0), "mpu_packed_buffer")
    with device.get_core(0):
        lhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 128], "row_major"),
            0,
            None,
            "sram",
        )
        rhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.integer("int8"), [8704], "mpu_packed_buffer"),
            0,
            None,
            "mpu_packed_buffer",
        )
        result = device.mm_weight_quant(lhs, rhs, legato.types.float("float8"), [128, 64], "per_output_channel")
        legato.tensor.memcpy(out, result)


@legato.compile(
    backend=lambda ctx: bertha.Bertha(ctx, "bertha"), param_kinds={"a": "universal", "w": "top", "out": "core"}
)
def mm_weight_quant_ic_128(
    a: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 128], "shared_memory"),
    w: legato.get_model().tensor_type(legato.types.integer("int8"), [8704], "shared_dram"),
    out: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 64], "shared_dram"),
):
    """Compile FP8 packed matmul with the default 128-channel group."""
    device = legato.get_context().get_device()
    with device.get_top():
        legato.tensor.request_load(a, device.get_core(0), "sram")
        legato.tensor.request_load(w, device.get_core(0), "mpu_packed_buffer")
    with device.get_core(0):
        lhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 128], "row_major"),
            0,
            None,
            "sram",
        )
        rhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.integer("int8"), [8704], "mpu_packed_buffer"),
            0,
            None,
            "mpu_packed_buffer",
        )
        result = device.mm_weight_quant(lhs, rhs, legato.types.float("float8"), [128, 64], "per_input_channel")
        legato.tensor.memcpy(out, result)


@legato.compile(
    backend=lambda ctx: bertha.Bertha(ctx, "bertha"), param_kinds={"a": "universal", "w": "top", "out": "core"}
)
def mm_weight_quant_ic_256(
    a: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 256], "shared_memory"),
    w: legato.get_model().tensor_type(legato.types.integer("int8"), [16896], "shared_dram"),
    out: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 64], "shared_dram"),
):
    """Compile FP8 packed matmul with an explicit 256-channel group."""
    device = legato.get_context().get_device()
    with device.get_top():
        legato.tensor.request_load(a, device.get_core(0), "sram")
        legato.tensor.request_load(w, device.get_core(0), "mpu_packed_buffer")
    with device.get_core(0):
        lhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 256], "row_major"),
            0,
            None,
            "sram",
        )
        rhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.integer("int8"), [16896], "mpu_packed_buffer"),
            0,
            None,
            "mpu_packed_buffer",
        )
        result = device.mm_weight_quant(lhs, rhs, legato.types.float("float8"), [256, 64], "per_input_channel", 256)
        legato.tensor.memcpy(out, result)


@legato.compile(
    backend=lambda ctx: bertha.Bertha(ctx, "bertha"), param_kinds={"a": "universal", "w": "top", "out": "core"}
)
def addmm_weight_quant_ic(
    a: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 128], "shared_memory"),
    w: legato.get_model().tensor_type(legato.types.integer("int8"), [9216], "shared_dram"),
    out: legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 64], "shared_dram"),
):
    """Compile an FP8 packed matmul with a packed FP32 bias."""
    device = legato.get_context().get_device()
    with device.get_top():
        legato.tensor.request_load(a, device.get_core(0), "sram")
        legato.tensor.request_load(w, device.get_core(0), "mpu_packed_buffer")
    with device.get_core(0):
        lhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.float("bfloat16"), [4, 128], "row_major"),
            0,
            None,
            "sram",
        )
        rhs = legato.tensor.receive_type(
            legato.get_model().tensor_type(legato.types.integer("int8"), [9216], "mpu_packed_buffer"),
            0,
            None,
            "mpu_packed_buffer",
        )
        result = device.addmm_weight_quant(
            lhs,
            rhs,
            legato.types.float("float8"),
            [128, 64],
            "per_input_channel",
            legato.types.float("float32"),
        )
        legato.tensor.memcpy(out, result)


CASES = {
    "mm_weight_quant_oc": (mm_weight_quant_oc, 128, QuantType.OutputChannelQuant, 128, False, 8704),
    "mm_weight_quant_ic_128": (mm_weight_quant_ic_128, 128, QuantType.InputChannelWiseQuant, 128, False, 8704),
    "mm_weight_quant_ic_256": (mm_weight_quant_ic_256, 256, QuantType.InputChannelWiseQuant, 256, False, 16896),
    "addmm_weight_quant_ic": (addmm_weight_quant_ic, 128, QuantType.InputChannelWiseQuant, 128, True, 9216),
}


def verify_case(name):
    """Check a packed weight's byte count and compile its kernel without launching."""
    torch.random.default_generator.manual_seed(0)
    kernel, channels, quant_type, group_size, use_bias, expected_bytes = CASES[name]
    # int8 is the transport container for real FP8 E4M3FN encodings, not integer quantization.
    weight = (torch.randn(channels, 64) * 0.25).to(torch.float8_e4m3fn).view(torch.int8)
    scale_shape = (64,) if quant_type == QuantType.OutputChannelQuant else (channels // group_size, 64)
    scale = torch.rand(scale_shape, dtype=torch.float32) * 0.09 + 0.01
    bias = torch.randn(64) * 0.1 if use_bias else None
    packed = pack_fp8_weight(weight, scale, quant_type=quant_type, group_size=group_size, bias_f32=bias)
    assert packed.numel() == expected_bytes, (kernel.__name__, packed.numel(), expected_bytes)
    activation = torch.randn(4, channels, dtype=torch.bfloat16)
    output = torch.empty(4, 64, dtype=torch.bfloat16)
    # Compile a binary without initializing a device or launching.
    with legato.session(output_type=legato.OutputType.BINARY, output_path=str(OUTPUT), launch=False):
        kernel(activation, legato.Arg(packed, bertha.top), legato.Arg(output, bertha.core(0)))


def main():
    """Check converter byte counts and lower all four variants to a binary artifact."""
    for name in CASES:
        verify_case(name)


if __name__ == "__main__":
    main()

The parameter annotation on w uses shared_dram storage; its receive type uses mpu_packed_buffer. Do not substitute mpu_weight, which expects a tiled rank-2-or-higher tensor. For other configurations, derive the packed size from hart.weight_converter.get_converted_size or the converter output. See Bertha operations and Bertha layouts.