Fetch a sharded model in one action, not one shard at a time
CI / test (push) Successful in 1m39s

The explorer listed every weight file with its own button. For a repository of
alternatives that is right — q4_k_m or q8_0, you want one — and for a sharded
model it is fifteen clicks to fetch one thing, with nothing stopping you from
losing count and leaving a version with fourteen of its parts.

Files are grouped into sets now. A run of "-00002-of-00004" is one row and one
button; everything else stays its own row.

Not "fetch all the weights", which would be wrong half the time: two safetensors
files are halves of one thing and useless apart, two GGUF files are usually
alternatives and fetching both spends gigabytes on a copy nobody asked for. The
set is the unit that is whole.

A repository can hold both the split and the joined copy of the same
quantisation — the real Qwen coder repos do — so a shard set keeps the -of-
notation in its name rather than collapsing onto the name of the whole file
beside it.

A set publishes no single checksum. Several files have several, and showing one
of them invites comparing it against the wrong thing.

The form posts every source in the set, and a failure part way says how many
already started: a set that fails on its fourth shard has three arriving, and
reading that as "nothing happened" sends somebody to fetch them all again.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5 committed 2026-09-27 15:32:27 +02:00
1 parent ed530133b6
commit a588646b7c
5 files changed
+277 -31

No files matched your search

+64
View File
@@ -17,6 +17,7 @@ import (
"io"
"net/http"
"net/url"
"regexp"
"strings"
"time"
)
@@ -257,3 +258,66 @@ func (c *Client) get(ctx context.Context, path string, out any) error {
}
return nil
}
// shardPattern matches the "-00002-of-00004" a split file carries before its
// extension. Five digits in practice, but not required: the count is what
// matters and a repository that used three would otherwise list every part as
// a separate model.
var shardPattern = regexp.MustCompile(`^(.*)-(\d+)-of-(\d+)(\.[^.]+)$`)
// FileSet is one thing worth fetching: a file, or every shard of one.
//
// The distinction the page turns on. Two GGUF files in a repository are usually
// alternatives — q4_k_m or q8_0, you want one — while two safetensors files are
// halves of the same weights and useless apart. Fetching "all the weights" is
// right for the second and wasteful for the first, so neither is offered:
// a set is offered, and a set is whole.
type FileSet struct {
// Name is what to call it. For shards it keeps the -of- notation, because
// a repository can hold both the split and the joined copy of the same
// quantisation and they must not read as one entry.
Name string
Files []File
Bytes int64
Shards int
}
// SHA256 is the checksum when there is one to give.
//
// Empty for a shard set: several files have several checksums and no single
// value for the set, and showing one of them invites comparing it against the
// wrong thing.
func (s FileSet) SHA256() string {
if len(s.Files) == 1 {
return s.Files[0].SHA256
}
return ""
}
// Group collects shards into sets, leaving every other file on its own.
//
// Order is preserved from the listing, so a repository that sorts its files
// sensibly still reads sensibly here.
func Group(files []File) []FileSet {
var out []FileSet
at := map[string]int{}
for _, f := range files {
m := shardPattern.FindStringSubmatch(f.Path)
if m == nil {
out = append(out, FileSet{Name: f.Path, Files: []File{f}, Bytes: f.Bytes})
continue
}
prefix, total, ext := m[1], m[3], m[4]
name := prefix + "-*-of-" + total + ext
i, ok := at[name]
if !ok {
at[name] = len(out)
out = append(out, FileSet{Name: name})
i = at[name]
}
out[i].Files = append(out[i].Files, f)
out[i].Bytes += f.Bytes
out[i].Shards = len(out[i].Files)
}
return out
}
+51
View File
@@ -136,3 +136,54 @@ func TestWeightsAreToldApartFromDocumentation(t *testing.T) {
}
}
}
func TestShardsBecomeOneSetAndAlternativesStaySeparate(t *testing.T) {
// Straight from a real repository, which holds both the split and the
// joined copy of the same quantisation.
sets := Group([]File{
{Path: "qwen-fp16-00001-of-00004.gguf", Bytes: 1},
{Path: "qwen-fp16-00002-of-00004.gguf", Bytes: 1},
{Path: "qwen-fp16-00003-of-00004.gguf", Bytes: 1},
{Path: "qwen-fp16-00004-of-00004.gguf", Bytes: 1},
{Path: "qwen-fp16.gguf", Bytes: 4},
{Path: "qwen-q4_k_m.gguf", Bytes: 2},
})
if len(sets) != 3 {
t.Fatalf("expected the shards as one and the two whole files apart, got %d: %+v",
len(sets), sets)
}
if sets[0].Shards != 4 || sets[0].Bytes != 4 {
t.Errorf("the shard set should total its parts, got %+v", sets[0])
}
// Two GGUFs in a repository are alternatives: fetching both is wasting
// gigabytes on a copy nobody asked for.
if sets[1].Shards != 0 || sets[2].Shards != 0 {
t.Errorf("whole files are not shard sets: %+v %+v", sets[1], sets[2])
}
}
func TestASplitAndAJoinedCopyDoNotReadAsOneEntry(t *testing.T) {
sets := Group([]File{
{Path: "m-q4-00001-of-00002.gguf"}, {Path: "m-q4-00002-of-00002.gguf"},
{Path: "m-q4.gguf"},
})
if sets[0].Name == sets[1].Name {
t.Errorf("the split and the joined copy collided on %q", sets[0].Name)
}
if !strings.Contains(sets[0].Name, "-of-00002") {
t.Errorf("a shard set should keep the notation; got %q", sets[0].Name)
}
}
func TestAShardSetPublishesNoSingleChecksum(t *testing.T) {
sets := Group([]File{
{Path: "m-00001-of-00002.safetensors", SHA256: "aaa"},
{Path: "m-00002-of-00002.safetensors", SHA256: "bbb"},
})
// Several files have several checksums and no single value for the set;
// showing one invites comparing it against the wrong thing.
if got := sets[0].SHA256(); got != "" {
t.Errorf("expected no checksum for a set, got %q", got)
}
}