mirror of
https://github.com/ollama/ollama.git
synced 2026-07-09 09:23:14 +00:00
964ea42c09
This is a rewrite of the create functionality for the MLX engine. The core idea behind the create functionality is to break the import/convert into a pipeline of distinct phases: * Read (scan the safetensors directory for the various bits of metadata) * Classify (determine what the import type) * Plan (determine any transforms that need to be done) * Write (transform any data as necessary and write out the blobs) * Create the manifest Each architecture has a "policy" which determines how to convert the model correctly. A number of different formats for safetensors are supported including: * nvfp4 (two formats: model optimized, torch) * fp8 datatypes (convert to mxfp8) * standard bf16 based weights A number of cleanups/simplifications have been done including: * using the baked in names for the tensors instead of munging them into something else * unified 3d expert tensors (instead of separate per expert tensors) * fewer unnecessary transforms to the various tensors in a model (keep a model as close to the source as possible) * unified capability checking * draft model handling (for MTP) is done on the same path Image generation has been intentionally removed.
58 lines
1.5 KiB
Go
58 lines
1.5 KiB
Go
package quant
|
|
|
|
import "testing"
|
|
|
|
func TestParams(t *testing.T) {
|
|
tests := []struct {
|
|
in string
|
|
groupSize int
|
|
bits int
|
|
mode string
|
|
}{
|
|
{"nvfp4", 16, 4, "nvfp4"},
|
|
{"NVFP4", 16, 4, "nvfp4"},
|
|
{"mxfp4", 32, 4, "mxfp4"},
|
|
{"mxfp8", 32, 8, "mxfp8"},
|
|
{"int4", 64, 4, "affine"},
|
|
{"int8", 64, 8, "affine"},
|
|
{"fp4", 64, 4, "affine"},
|
|
{"q4", 64, 4, "affine"},
|
|
{"fp8", 64, 8, "affine"},
|
|
{"q8", 64, 8, "affine"},
|
|
{"", 0, 0, ""},
|
|
// Unknown non-empty types fall back to 8-bit affine, matching the
|
|
// runtime loader's historical behavior.
|
|
{"something-else", 32, 8, "affine"},
|
|
}
|
|
for _, tt := range tests {
|
|
gs, bits, mode := Params(tt.in)
|
|
if gs != tt.groupSize || bits != tt.bits || mode != tt.mode {
|
|
t.Errorf("Params(%q) = (%d, %d, %q), want (%d, %d, %q)", tt.in, gs, bits, mode, tt.groupSize, tt.bits, tt.mode)
|
|
}
|
|
}
|
|
}
|
|
|
|
func TestBitsAndPackFactor(t *testing.T) {
|
|
tests := []struct {
|
|
in string
|
|
bits int
|
|
packFactor int
|
|
}{
|
|
{"int4", 4, 8},
|
|
{"nvfp4", 4, 8},
|
|
{"mxfp4", 4, 8}, // regression: mxfp4 was missing from the old show.go switch
|
|
{"int8", 8, 4},
|
|
{"mxfp8", 8, 4},
|
|
{"FP8", 8, 4},
|
|
{"", 0, 0},
|
|
{"unknown", 0, 0}, // strict: no fallback for sizing/display
|
|
}
|
|
for _, tt := range tests {
|
|
if got := Bits(tt.in); got != tt.bits {
|
|
t.Errorf("Bits(%q) = %d, want %d", tt.in, got, tt.bits)
|
|
}
|
|
if got := PackFactor(tt.in); got != tt.packFactor {
|
|
t.Errorf("PackFactor(%q) = %d, want %d", tt.in, got, tt.packFactor)
|
|
}
|
|
}
|
|
}
|