llmgw/internal/web/admin_mirror.go

package web

import (
	"context"
	"errors"
	"fmt"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
	"gitea.mynodes.xyz/ilker/llmgw/internal/gguf"
	"html/template"
	"net/http"
	"sort"
	"strconv"
	"strings"
	"time"

	"gitea.mynodes.xyz/ilker/llmgw/internal/hf"
	"gitea.mynodes.xyz/ilker/llmgw/internal/mirror"
	"gitea.mynodes.xyz/ilker/llmgw/internal/sourcecheck"
	"gitea.mynodes.xyz/ilker/llmgw/internal/store"
)

// MirrorAdmin is the mirror's management API as this page needs it.
//
// An interface rather than the client itself so the page can be tested without
// a mirror, and so a deployment without one leaves it nil and the page
// unmounted — the same shape the cluster dashboard uses.
type MirrorAdmin interface {
	List(ctx context.Context) (mirror.ModelsResponse, error)
	Fetch(ctx context.Context, req mirror.FetchRequest) (mirror.FetchResponse, error)
	Remove(ctx context.Context, model, version string) ([]string, error)
	// Addr is the mirror this points at, shown on the page because an operator
	// looking at an empty table needs to know which machine is empty.
	Addr() string
}

// Registry is the model registry the explorer browses. Nil leaves the explorer
// off, which is right for a deployment with no way out to the internet: a
// search box that can only time out is worse than no search box.
type Registry interface {
	Explore(ctx context.Context, q hf.Query) ([]hf.Model, error)
	Repo(ctx context.Context, id string) (hf.Repo, error)
}

// repoSet is one thing worth fetching: a file, or every shard of one.
type repoSet struct {
	// Verdict is whether the nodes' llama.cpp can load it, from the
	// architecture the registry reports; a set that cannot be served is
	// shown with the reason and no fetch button.
	Verdict gguf.Verdict
	hf.FileSet
	// Sources are every file in the set, in the order the registry listed them.
	// A shard set goes as one action — a fifteen-shard model is not fifteen
	// decisions, it is one.
	Sources []repoSource
	// Model and Version are what the fetch form starts with. Editable, because
	// the name a catalog entry uses is the operator's to choose and the
	// registry's idea of it is only a good guess.
	Model   string
	Version string
}

type repoSource struct{ URL, Path string }

// modelUse is a catalog model that runs a mirrored version, and how.
type modelUse struct {
	Name     string
	Mode     string // always, ondemand or stopped
	Replicas int
	// Declared models are read from server.yaml on every start, so a mode
	// changed here would revert; the page says where to change it instead.
	Declared bool
}

// mirrorRow is one model version as the page shows it.
type mirrorRow struct {
	mirror.Status
	// URL is what a catalog entry should say. The one string on this page
	// somebody will copy, so it is rendered rather than left to be assembled.
	URL     string
	Percent int
	// Serve is the offer to put this version into service, open on the row the
	// operator clicked. Nil on a row that is still arriving: a catalog entry
	// pointing at a partial file is the thing versioned paths exist to prevent.
	Serve *serveForm

	// HasPercent is false when the upstream never sent a length. The bar is
	// then not drawn at all: a progress bar that has to invent its own total is
	// worse than a byte count that only goes up.
	HasPercent bool

	// Files is how many files a held version has. A version of several is one
	// row — a fifteen-shard model is one thing to serve or remove, and fifteen
	// rows of it pushed everything else off the page.
	Files int
	// UsedBy is the catalog models whose source is this row's address, and how
	// each runs, so a file in use does not look like one waiting to be served
	// and how it runs can be changed where it is seen.
	UsedBy []modelUse
}

type adminMirrorData struct {
	Nav  Nav
	Addr string
	Base string
	Rows []mirrorRow
	// Groups are the same rows under their model.
	Groups []mirrorGroup

	Fetching int
	Present  int
	Failed   int

	// Unreachable separates "the mirror holds nothing" from "the mirror did not
	// answer". They render almost identically and mean opposite things: one is
	// a mirror waiting to be filled, the other is a machine to go and look at.
	Unreachable bool

	// HasCatalog is false where the catalogue cannot be changed at runtime, and
	// the serve action is then not offered rather than offered and refused.
	HasCatalog bool

	// The explorer.
	HasRegistry bool
	Query       string
	Explore     *explorer
	// Back is the listing an open repository was reached from.
	Back template.URL
	// Inline is a repository rendered beneath its row in the listing, without
	// the page around it.
	Inline bool
	// Card is the open repository's model card, rendered.
	Card template.HTML
	// RepoURL is the open repository on the registry's own site.
	RepoURL     string
	Repo        *hf.Repo
	RepoSets    []repoSet
	RegistryErr string

	Message string
	Error   string
	CSRF    string
	Version string
	Now     string
	// RefreshSec is 0 when nothing is downloading. The page reloads only while
	// there is something to watch, because a reload while an operator is typing
	// a source URL into the form takes the URL with it.
	RefreshSec int
}

// RegisterAdminMirror mounts the model mirror under /admin/mirror.
//
// Fetching and removing, and nothing else. What the mirror holds is operator
// state rather than account state, but the reasoning is the same as
// /admin/users: these two are the work somebody does repeatedly and should not
// need a shell for, while anything that would let a stolen session reach
// further — pointing the control plane at a different mirror, say — stays in
// configuration where changing it means access to the host.
// MirrorDeps is what the mirror page can reach.
//
// A struct rather than five parameters: three of them are optional interfaces
// and two are the same shape, which at a call site is a swap nobody notices
// until a page shows the wrong thing.
type MirrorDeps struct {
	Mirror MirrorAdmin
	// Registry browses a model registry. Nil turns the explorer off.
	Registry Registry
	// Catalog puts a mirrored model into service. Nil leaves the serve action
	// off, which is right where the catalogue cannot be changed at runtime.
	Catalog CatalogAdmin
	// Fleet is the cluster, for saying whether anything could run a model
	// before somebody spends an hour downloading it.
	Fleet Source
}

func (p *AccountPages) RegisterAdminMirror(mux *http.ServeMux, d MirrorDeps) {
	m := d.Mirror
	p.hasMirror = true
	p.registry = d.Registry
	p.catalog = d.Catalog
	p.fleet = d.Fleet

	mux.HandleFunc("GET /admin/mirror", p.adminOnly(
		func(w http.ResponseWriter, r *http.Request, _ store.User) {
			p.renderMirror(w, r, m, "", "")
		}))

	// One repository as a fragment, for the row beneath it in the listing.
	mux.HandleFunc("GET /admin/mirror/repo", p.adminOnly(
		func(w http.ResponseWriter, r *http.Request, _ store.User) {
			repo := strings.TrimSpace(r.URL.Query().Get("repo"))
			if p.registry == nil || repo == "" {
				http.NotFound(w, r)
				return
			}
			d := adminMirrorData{CSRF: csrfFor(r), Inline: true, HasRegistry: true}
			p.openRepo(r, &d, repo)
			if d.RegistryErr != "" {
				http.Error(w, "The registry: "+d.RegistryErr, http.StatusBadGateway)
				return
			}
			render(w, "mirror_repo", d)
		}))

	mux.HandleFunc("POST /admin/mirror/fetch", p.mirrorAction(m,
		func(ctx context.Context, admin store.User, r *http.Request) (string, error) {
			// One form posts one source or fifteen. A sharded model is one
			// decision, not one per shard, and the mirror fetches each file
			// independently anyway.
			sources, files := r.Form["source"], r.Form["file"]
			if len(sources) == 0 || strings.TrimSpace(sources[0]) == "" {
				return "", errors.New("a source URL is required")
			}
			model := strings.TrimSpace(r.FormValue("model"))
			version := strings.TrimSpace(r.FormValue("version"))

			// Read before anything is downloaded: the header says whether the
			// nodes can load it, and a file that cannot be served is an hour
			// of download and a replica failing on every node.
			if r.FormValue("force") == "" {
				if v := p.checkSource(ctx, strings.TrimSpace(sources[0])); !v.Servable {
					return "", fmt.Errorf("not fetched: %s. Tick \"fetch anyway\" to take it regardless", v.Reason)
				}
			}

			var started int
			for i, src := range sources {
				src = strings.TrimSpace(src)
				if src == "" {
					continue
				}
				req := mirror.FetchRequest{Source: src, Model: model, Version: version}
				if i < len(files) {
					req.File = strings.TrimSpace(files[i])
				}
				if _, err := m.Fetch(ctx, req); err != nil {
					// Reported with what already went, so a set that fails on
					// its fourth shard does not read as though nothing started.
					return "", fmt.Errorf("%d of %d started, then %s: %w",
						started, len(sources), mirror.LastPathElement(src), err)
				}
				started++
			}
			// Named as started, not as done. The mirror answered 202 and the
			// bytes are still moving; a message saying "added" would invite
			// somebody to point the catalog at a partial file.
			return fmt.Sprintf("started downloading %s/%s (%s)",
				model, version, plural(started, "file", "files")), nil
		}))

	if d.Catalog != nil {
		// How a model runs, changed where the file it runs is seen.
		mux.HandleFunc("POST /admin/mirror/mode", p.mirrorAction(m,
			func(ctx context.Context, admin store.User, r *http.Request) (string, error) {
				name, mode := r.FormValue("model"), r.FormValue("mode")
				if name == "" || (mode != modeAlways && mode != modeOnDemand && mode != modeStopped) {
					return "", errors.New("choose a model and how it should run")
				}
				replicas, _ := strconv.Atoi(r.FormValue("replicas"))
				sc := scalingFor(mode, replicas, 0)
				if _, err := d.Catalog.Scale(ctx, name, sc, admin.Email); err != nil {
					return "", err
				}
				switch mode {
				case modeOnDemand:
					return name + " now loads when asked for and stops after 15 minutes unused", nil
				case modeStopped:
					return name + " is stopped; its replicas are being taken down", nil
				}
				return name + " now runs all the time", nil
			}))

		mux.HandleFunc("POST /admin/mirror/serve", p.mirrorAction(m,
			func(ctx context.Context, admin store.User, r *http.Request) (string, error) {
				model, err := buildModel(r.FormValue)
				if err != nil {
					return "", err
				}
				// The same check `llmgw ctl model add` makes. This page builds
				// the source out of a row the mirror reported, so it is right
				// far more often than something typed — but a file removed
				// between the listing and the click, or a mirror that has gone
				// down since, both end as a replica failing on a node minutes
				// later with an error naming the node. Asking costs one HEAD.
				if _, err := sourcecheck.Check(ctx, model.Source); err != nil {
					return "", err
				}
				if err := d.Catalog.Add(ctx, model, admin.Email); err != nil {
					return "", err
				}
				// Placed on the next pass, not now: saying it is serving would
				// be wrong for however long the weights take to reach a node.
				return fmt.Sprintf(
					"%s added to the catalog, %s; it is placed on the next pass",
					model.Name, plural(model.Scaling.Min, "replica", "replicas")), nil
			}))
	}

	mux.HandleFunc("POST /admin/mirror/{model}/{version}/remove", p.mirrorAction(m,
		func(ctx context.Context, _ store.User, r *http.Request) (string, error) {
			model, ver := r.PathValue("model"), r.PathValue("version")
			removed, err := m.Remove(ctx, model, ver)
			if err != nil {
				return "", err
			}
			return "removed " + model + "/" + ver +
				" (" + plural(len(removed), "file", "files") + ")", nil
		}))
}

// mirrorAction wraps a change: admin session, CSRF, audit line, re-render.
//
// The audit line is not optional. Removing a version deletes weights every node
// may be about to fetch, and the only record that it happened on purpose is
// this log entry naming who clicked it.
func (p *AccountPages) mirrorAction(m MirrorAdmin,
	do func(ctx context.Context, admin store.User, r *http.Request) (string, error),
) http.HandlerFunc {
	return p.adminOnly(func(w http.ResponseWriter, r *http.Request, admin store.User) {
		sess, _, _ := p.session(r)
		if !p.accounts.CheckCSRF(sess, r.FormValue("csrf")) {
			http.Error(w, "bad request", http.StatusBadRequest)
			return
		}
		msg, err := do(r.Context(), admin, r)
		if err != nil {
			p.log.Warn("a mirror action was refused",
				"admin", admin.Email, "mirror", m.Addr(), "err", err)
			p.renderMirror(w, r, m, "", err.Error())
			return
		}
		p.log.Info("the mirror was changed",
			"admin", admin.Email, "mirror", m.Addr(), "what", msg)
		p.renderMirror(w, r, m, msg, "")
	})
}

func (p *AccountPages) renderMirror(w http.ResponseWriter, r *http.Request,
	m MirrorAdmin, msg, errMsg string,
) {
	d := adminMirrorData{
		Nav:     navFrom(r, "mirror"),
		Addr:    m.Addr(),
		Message: msg,
		Error:   errMsg,
		CSRF:    csrfFor(r),
		Version: p.version,
		Now:     time.Now().UTC().Format(time.RFC3339),
	}

	d.HasCatalog = p.catalog != nil
	p.explore(r, &d)

	got, err := m.List(r.Context())
	if err != nil {
		p.log.Error("could not read the mirror", "mirror", m.Addr(), "err", err)
		d.Unreachable = true
		if d.Error == "" {
			d.Error = err.Error()
		}
		render(w, "admin_mirror", d)
		return
	}

	d.Base = got.Base
	var sources map[string][]catalog.Model
	if p.catalog != nil {
		if sources, err = p.catalog.Sources(r.Context()); err != nil {
			p.log.Warn("could not read which models use the mirror", "err", err)
		}
	}
	held := map[string]int{} // model/version -> index of its row
	for _, st := range got.Models {
		switch st.State {
		case mirror.Fetching:
			d.Fetching++
		case mirror.Failed:
			d.Failed++
		default:
			key := st.Model + "/" + st.Version
			if i, ok := held[key]; ok {
				// Another file of a version already shown: counted into it.
				d.Rows[i].Files++
				d.Rows[i].BytesDone += st.BytesDone
				d.Rows[i].URL = strings.TrimSuffix(
					mirror.Entry{Model: st.Model, Version: st.Version}.URL(got.Base), "/") + "/"
				continue
			}
			held[key] = len(d.Rows)
			d.Present++ // versions, not files: what the rows below count
		}
		row := mirrorRow{
			Status: st,
			Files:  1,
			URL: mirror.Entry{Model: st.Model, Version: st.Version, File: st.File}.
				URL(got.Base),
		}
		if pc, ok := st.Percent(); ok {
			row.Percent, row.HasPercent = pc, true
		}
		d.Rows = append(d.Rows, row)
	}
	for i := range d.Rows {
		for _, m := range sources[d.Rows[i].URL] {
			d.Rows[i].UsedBy = append(d.Rows[i].UsedBy, modelUse{
				Name: m.Name, Mode: modeOf(m.Scaling), Replicas: max(m.Scaling.Max, 1),
				Declared: p.catalog.Declared(m.Name),
			})
		}
	}

	// The one row whose serve form is open, and only when the catalog can be
	// changed at all.
	if p.catalog != nil {
		if want := r.URL.Query().Get("serve"); want != "" {
			byVersion := map[string][]mirror.Status{}
			for _, st := range got.Models {
				byVersion[st.Model+"/"+st.Version] = append(byVersion[st.Model+"/"+st.Version], st)
			}
			for i := range d.Rows {
				key := d.Rows[i].Model + ":" + d.Rows[i].Version
				if key != want || d.Rows[i].State != mirror.Present {
					continue
				}
				form := planServe(d.Rows[i].Status,
					byVersion[d.Rows[i].Model+"/"+d.Rows[i].Version], got.Base, p.fleet, p.catalog, func(src string) gguf.Verdict { return p.checkSource(r.Context(), src) })
				d.Rows[i].Serve = &form
				break
			}
		}
	}

	// Downloads first, then failures, then what is simply here. The two states
	// somebody opened this page for are at the top; the rest is a list they
	// scroll. Within a state the order is fixed so rows do not shuffle between
	// reloads while a byte count ticks.
	sort.SliceStable(d.Rows, func(i, j int) bool {
		a, b := d.Rows[i], d.Rows[j]
		if ra, rb := mirrorRank(a.State), mirrorRank(b.State); ra != rb {
			return ra < rb
		}
		if a.Model != b.Model {
			return a.Model < b.Model
		}
		if a.Version != b.Version {
			return a.Version < b.Version
		}
		return a.File < b.File
	})

	d.Groups = groupByModel(d.Rows)

	if d.Fetching > 0 {
		d.RefreshSec = 5
	}
	render(w, "admin_mirror", d)
}

// mirrorGroup is a family of models the mirror holds — every qwen, every
// llama — and their versions.
type mirrorGroup struct {
	Family string
	Rows   []mirrorRow
	// Models is how many different models are in it; a family of one is
	// shown without a heading, which would only repeat its row.
	Models int
	// Bytes is what the versions take on disk, downloads so far included.
	Bytes            int64
	Fetching, Failed int
	InUse            bool
}

// modelFamily is the family a model's name puts it in: the letters its name
// starts with. "qwen2.5-coder-7b-instruct" and "qwen3-8b" are qwen,
// "llama-3.1-8b-instruct" is llama, "deepseek-r1-distill-qwen-14b" is
// deepseek — which is how people talk about them, and all a name says.
func modelFamily(model string) string {
	name := strings.ToLower(model)
	end := strings.IndexFunc(name, func(r rune) bool { return r < 'a' || r > 'z' })
	if end <= 0 {
		// Starts with a digit or a symbol: the first part whole.
		first, _, _ := strings.Cut(name, "-")
		return first
	}
	return name[:end]
}

// groupByModel gathers rows by family, in the order the rows are already in:
// a family with a download or a failure comes first because its first row
// does, and its models and versions stay in that order beneath it.
//
// Grouped because a family is what somebody looks for. Sorted by state and
// name alone, "which qwens do we have" meant reading the whole table.
func groupByModel(rows []mirrorRow) []mirrorGroup {
	var out []mirrorGroup
	at := map[string]int{}
	models := map[string]map[string]bool{}
	for _, r := range rows {
		fam := modelFamily(r.Model)
		i, ok := at[fam]
		if !ok {
			i = len(out)
			at[fam] = i
			out = append(out, mirrorGroup{Family: fam})
			models[fam] = map[string]bool{}
		}
		g := &out[i]
		g.Rows = append(g.Rows, r)
		g.Bytes += r.BytesDone
		models[fam][r.Model] = true
		switch r.State {
		case mirror.Fetching:
			g.Fetching++
		case mirror.Failed:
			g.Failed++
		}
		if len(r.UsedBy) > 0 {
			g.InUse = true
		}
	}
	for i := range out {
		out[i].Models = len(models[out[i].Family])
	}
	return out
}

func mirrorRank(s mirror.State) int {
	switch s {
	case mirror.Fetching:
		return 0
	case mirror.Failed:
		return 1
	default:
		return 2
	}
}

func plural(n int, one, many string) string {
	if n == 1 {
		return "1 " + one
	}
	return strconv.Itoa(n) + " " + many
}

// explore fills in the registry half of the page: a listing, or one repository.
//
// A failure here is reported in its own line rather than as the page's error.
// The registry is somebody else's server on the other side of the internet, and
// it being down says nothing about the mirror — which is the thing this page is
// actually for, and which must keep working when Hugging Face does not.
func (p *AccountPages) explore(r *http.Request, d *adminMirrorData) {
	if p.registry == nil {
		return
	}
	d.HasRegistry = true
	d.Query = strings.TrimSpace(r.URL.Query().Get("q"))

	if repo := strings.TrimSpace(r.URL.Query().Get("repo")); repo != "" {
		d.Back = exploreState(r.URL.Query()).href(d.Query, "", "")
		p.openRepo(r, d, repo)
		return
	}

	p.exploreList(r, d)
}

// openRepo reads one repository: its card, and its weights grouped into what
// can be fetched.
func (p *AccountPages) openRepo(r *http.Request, d *adminMirrorData, repo string) {
	got, err := p.registry.Repo(r.Context(), repo)
	if err != nil {
		d.RegistryErr = err.Error()
		return
	}
	d.Repo = &got
	d.RepoURL = strings.TrimSuffix(got.DownloadURL("", ""), "/resolve/"+revOrMain(got.Revision)+"/")
	d.Card = renderCard(got.Card, d.RepoURL+"/blob/"+revOrMain(got.Revision)+"/")
	var weights []hf.File
	for _, f := range got.Files {
		if !f.IsWeights() {
			// Thirty files of which four are worth gigabytes. The rest are
			// a README and a tokenizer, and listing them buries the ones
			// somebody came for.
			continue
		}
		weights = append(weights, f)
	}
	for _, set := range hf.Group(weights) {
		rs := repoSet{
			FileSet: set,
			Model:   defaultModelName(got.ID),
			// The revision, not the date: two fetches of the same commit
			// land in the same place, and a repository re-uploaded under
			// the same branch lands somewhere else.
			Version: got.ShortRevision(),
		}
		rs.Verdict = p.verdictFor(got.Architecture, set.Name)
		for _, f := range set.Files {
			rs.Sources = append(rs.Sources,
				repoSource{URL: got.DownloadURL("", f.Path), Path: f.Path})
		}
		d.RepoSets = append(d.RepoSets, rs)
	}
}

// defaultModelName is the catalog name a repository suggests.
//
// The last element, lowercased, with the quantisation marker dropped: the owner
// is not part of what anybody calls the model, and "-GGUF" describes the file
// format rather than the model, which is already clear from the file being
// fetched.
func defaultModelName(repoID string) string {
	name := mirror.LastPathElement(repoID)
	name = strings.TrimSuffix(name, "-GGUF")
	name = strings.TrimSuffix(name, "-gguf")
	return strings.ToLower(name)
}

// verdictFor judges a registry file set by the repository's architecture. A
// vision projector is its own architecture whatever the repository says.
func (p *AccountPages) verdictFor(arch, name string) gguf.Verdict {
	if strings.Contains(strings.ToLower(name), "mmproj") {
		return gguf.LlamaCpp(p.llamaVersion, "clip")
	}
	// Published beside the quantisations they were made with, and a GGUF
	// file like them, so the repository's architecture would pass it.
	if strings.Contains(strings.ToLower(name), "imatrix") {
		return gguf.Verdict{Checked: true,
			Reason: "it is an importance matrix for quantising, not a model"}
	}
	if arch == "" || !gguf.IsGGUF(name) {
		return gguf.Verdict{Servable: true}
	}
	return gguf.LlamaCpp(p.llamaVersion, arch)
}

// checkSource reads a GGUF source's header and judges it; anything else, or a
// header that cannot be read, passes.
func (p *AccountPages) checkSource(ctx context.Context, source string) gguf.Verdict {
	url, ok := sourcecheck.HTTPSource(source)
	if !ok || !gguf.IsGGUF(url) || p.llamaVersion == "" {
		return gguf.Verdict{Servable: true}
	}
	return gguf.Check(ctx, url, p.llamaVersion)
}

func revOrMain(rev string) string {
	if rev == "" {
		return "main"
	}
	return rev
}