Introduce new routing backend (#790)
This is a huge backend change that essentially started with rewriting the concurrency handling for processes and blew up to a refactor of the entire application. In short these are the improvements: **Better state and life cycle management:** Life cycle management of processes has always been the trickiest part of the code. Juggling mutex locks between multiple locations to reduce race conditions was complex. Too complex for my feeble brain to build a simple mental model around as llama-swap gained more features. All of that has been refactored. Most of the locks are gone, replaced with a single run() that owns all state changes. There is one place to start from now to understand and extend routing logic. The improved life cycle management makes it easier to implement more complex swap optimization strategies in the future like #727. **Collation of requests:** llama-swap previously handled requests and swapping in the order they came in. For example requests for models in this order ABCABC would result in 5 swaps. Now those requests are handled in this order AABBCC. The result is less time waiting for swap under a high churn request queue. This fixes #588 #612. A possible future enhancement is to support a starvation parameter so swap can be forced when models have been waiting too long. **Shared base implementation for groups and swap matrix:** During the refactor it became clear that much of the swapping logic was shared between these two implementations. That is not surprising considering the swap matrix was added many moons after groups. Now they share a common base and their specific swap strategies are implemented into the swapPlanner interface. Requests for bespoke or specific swapping scenarios is a common theme in the issues. Now users can implement whatever bespoke and weird swapping strategy they want in their own fork. Just ask your agent of choice to implement swapPlanner. I'll still remaining more conservative on what actually lands in core llama-swap and will continue to evaluate PRs if the changes is good for everyone or just one specific use case. **AI / Agentic Disclosure:** I paid very close attention to the low level swap concurrency design and implementation. It's important to keep that essential part reliable, boring and no surprises. Backwards compatibility was also maintained, even the one way non-exclusive group model loading behaviour that people have rightly pointed out be a weird design decision. With the underlying swap core done the web server, api and UI sitting on top were largely ported over with Claude Code and Opus 4.7 in multiple phases. If you're curious I kept the changes in docs/newrouter-todo.md. I did several passes to make sure things weren't left behind. However, even frontier LLMs at the time of this PR still make small decisions that don't make a lot of sense. They get shit wrong all the time, just in small subtle way. That said, there's likely to be some new bugs introduced with this massive refactor. I'm fairly confident that there's no major architectural flaws that would cause goal seeking agents to make dumb, ugly code decisions. For a little while the legacy llama-swap will be available under cmd/legacy/llama-swap. The plan is to eventually delete that entry point as well as the proxy package. On a bit of a personal note, this PR is exciting and a bit sad for me. I hand wrote much of the original code and this PR ultimately replaces much of it. While the old code served as a good reference for the agent to implement the new stuff it still a bit sad to eventually delete it all.
This commit is contained in:
Benson Wong
GitHub
@@ -0,0 +1,376 @@
|
||||
package config
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
"sort"
|
||||
"strings"
|
||||
"unicode"
|
||||
)
|
||||
|
||||
const maxDSLExpansions = 1000
|
||||
|
||||
// Token types for the DSL lexer
|
||||
type tokenType int
|
||||
|
||||
const (
|
||||
tokIdent tokenType = iota // model alias or name
|
||||
tokAnd // &
|
||||
tokOr // |
|
||||
tokLParen // (
|
||||
tokRParen // )
|
||||
tokRef // +setName
|
||||
tokEOF
|
||||
)
|
||||
|
||||
type token struct {
|
||||
typ tokenType
|
||||
val string
|
||||
}
|
||||
|
||||
// tokenize splits a DSL string into tokens.
|
||||
func tokenize(input string) ([]token, error) {
|
||||
var tokens []token
|
||||
i := 0
|
||||
runes := []rune(input)
|
||||
|
||||
for i < len(runes) {
|
||||
ch := runes[i]
|
||||
|
||||
// skip whitespace
|
||||
if unicode.IsSpace(ch) {
|
||||
i++
|
||||
continue
|
||||
}
|
||||
|
||||
switch ch {
|
||||
case '&':
|
||||
tokens = append(tokens, token{tokAnd, "&"})
|
||||
i++
|
||||
case '|':
|
||||
tokens = append(tokens, token{tokOr, "|"})
|
||||
i++
|
||||
case '(':
|
||||
tokens = append(tokens, token{tokLParen, "("})
|
||||
i++
|
||||
case ')':
|
||||
tokens = append(tokens, token{tokRParen, ")"})
|
||||
i++
|
||||
case '+':
|
||||
// +ref: read the identifier that follows
|
||||
i++
|
||||
start := i
|
||||
for i < len(runes) && isIdentChar(runes[i]) {
|
||||
i++
|
||||
}
|
||||
if i == start {
|
||||
return nil, fmt.Errorf("expected set name after '+' at position %d", start)
|
||||
}
|
||||
tokens = append(tokens, token{tokRef, string(runes[start:i])})
|
||||
default:
|
||||
if isIdentChar(ch) {
|
||||
start := i
|
||||
for i < len(runes) && isIdentChar(runes[i]) {
|
||||
i++
|
||||
}
|
||||
tokens = append(tokens, token{tokIdent, string(runes[start:i])})
|
||||
} else {
|
||||
return nil, fmt.Errorf("unexpected character %q at position %d", ch, i)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
tokens = append(tokens, token{tokEOF, ""})
|
||||
return tokens, nil
|
||||
}
|
||||
|
||||
func isIdentChar(ch rune) bool {
|
||||
return unicode.IsLetter(ch) || unicode.IsDigit(ch) || ch == '_' || ch == '-' || ch == '.'
|
||||
}
|
||||
|
||||
// AST node types
|
||||
type dslNode interface {
|
||||
dslNode()
|
||||
}
|
||||
|
||||
type andNode struct {
|
||||
children []dslNode
|
||||
}
|
||||
|
||||
type orNode struct {
|
||||
children []dslNode
|
||||
}
|
||||
|
||||
type leafNode struct {
|
||||
name string
|
||||
}
|
||||
|
||||
type refNode struct {
|
||||
setName string
|
||||
}
|
||||
|
||||
func (andNode) dslNode() {}
|
||||
func (orNode) dslNode() {}
|
||||
func (leafNode) dslNode() {}
|
||||
func (refNode) dslNode() {}
|
||||
|
||||
// parser holds state for recursive-descent parsing.
|
||||
type parser struct {
|
||||
tokens []token
|
||||
pos int
|
||||
}
|
||||
|
||||
func (p *parser) peek() token {
|
||||
if p.pos < len(p.tokens) {
|
||||
return p.tokens[p.pos]
|
||||
}
|
||||
return token{tokEOF, ""}
|
||||
}
|
||||
|
||||
func (p *parser) next() token {
|
||||
t := p.peek()
|
||||
if t.typ != tokEOF {
|
||||
p.pos++
|
||||
}
|
||||
return t
|
||||
}
|
||||
|
||||
func (p *parser) expect(typ tokenType) (token, error) {
|
||||
t := p.next()
|
||||
if t.typ != typ {
|
||||
return t, fmt.Errorf("expected token type %d, got %q", typ, t.val)
|
||||
}
|
||||
return t, nil
|
||||
}
|
||||
|
||||
// Grammar:
|
||||
//
|
||||
// expr = andExpr
|
||||
// andExpr = orExpr ('&' orExpr)*
|
||||
// orExpr = atom ('|' atom)*
|
||||
// atom = ident | '+' ident | '(' expr ')'
|
||||
//
|
||||
// & binds tighter than |, so "a | b & c" means "a | (b & c)"
|
||||
func parse(tokens []token) (dslNode, error) {
|
||||
p := &parser{tokens: tokens}
|
||||
node, err := p.parseExpr()
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
if p.peek().typ != tokEOF {
|
||||
return nil, fmt.Errorf("unexpected token %q after expression", p.peek().val)
|
||||
}
|
||||
return node, nil
|
||||
}
|
||||
|
||||
func (p *parser) parseExpr() (dslNode, error) {
|
||||
return p.parseOrExpr()
|
||||
}
|
||||
|
||||
func (p *parser) parseOrExpr() (dslNode, error) {
|
||||
left, err := p.parseAndExpr()
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
if p.peek().typ == tokOr {
|
||||
children := []dslNode{left}
|
||||
for p.peek().typ == tokOr {
|
||||
p.next() // consume |
|
||||
right, err := p.parseAndExpr()
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
children = append(children, right)
|
||||
}
|
||||
return orNode{children: children}, nil
|
||||
}
|
||||
|
||||
return left, nil
|
||||
}
|
||||
|
||||
func (p *parser) parseAndExpr() (dslNode, error) {
|
||||
left, err := p.parseAtom()
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
if p.peek().typ == tokAnd {
|
||||
children := []dslNode{left}
|
||||
for p.peek().typ == tokAnd {
|
||||
p.next() // consume &
|
||||
right, err := p.parseAtom()
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
children = append(children, right)
|
||||
}
|
||||
return andNode{children: children}, nil
|
||||
}
|
||||
|
||||
return left, nil
|
||||
}
|
||||
|
||||
func (p *parser) parseAtom() (dslNode, error) {
|
||||
t := p.peek()
|
||||
|
||||
switch t.typ {
|
||||
case tokIdent:
|
||||
p.next()
|
||||
return leafNode{name: t.val}, nil
|
||||
|
||||
case tokRef:
|
||||
p.next()
|
||||
return refNode{setName: t.val}, nil
|
||||
|
||||
case tokLParen:
|
||||
p.next() // consume (
|
||||
node, err := p.parseExpr()
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
if _, err := p.expect(tokRParen); err != nil {
|
||||
return nil, fmt.Errorf("missing closing parenthesis")
|
||||
}
|
||||
return node, nil
|
||||
|
||||
default:
|
||||
return nil, fmt.Errorf("unexpected token %q", t.val)
|
||||
}
|
||||
}
|
||||
|
||||
// expand walks the AST and produces all combinations.
|
||||
// resolvedRefs contains previously expanded sets for +ref resolution.
|
||||
func expand(node dslNode, resolvedRefs map[string][][]string) ([][]string, error) {
|
||||
switch n := node.(type) {
|
||||
case leafNode:
|
||||
return [][]string{{n.name}}, nil
|
||||
|
||||
case refNode:
|
||||
expanded, ok := resolvedRefs[n.setName]
|
||||
if !ok {
|
||||
return nil, fmt.Errorf("unknown set reference +%s", n.setName)
|
||||
}
|
||||
// Return a copy
|
||||
result := make([][]string, len(expanded))
|
||||
for i, combo := range expanded {
|
||||
result[i] = make([]string, len(combo))
|
||||
copy(result[i], combo)
|
||||
}
|
||||
return result, nil
|
||||
|
||||
case orNode:
|
||||
// Union of all children's expansions
|
||||
var result [][]string
|
||||
for _, child := range n.children {
|
||||
childResult, err := expand(child, resolvedRefs)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
result = append(result, childResult...)
|
||||
if len(result) > maxDSLExpansions {
|
||||
return nil, fmt.Errorf("DSL expansion exceeded %d combinations", maxDSLExpansions)
|
||||
}
|
||||
}
|
||||
return result, nil
|
||||
|
||||
case andNode:
|
||||
// Cartesian product across children
|
||||
result := [][]string{{}} // start with one empty combo
|
||||
for _, child := range n.children {
|
||||
childResult, err := expand(child, resolvedRefs)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
result, err = cartesianProduct(result, childResult, maxDSLExpansions)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
}
|
||||
return result, nil
|
||||
|
||||
default:
|
||||
return nil, fmt.Errorf("unknown node type %T", node)
|
||||
}
|
||||
}
|
||||
|
||||
// cartesianProduct computes the cartesian product of two sets of combinations.
|
||||
// It returns an error if the product would exceed cap.
|
||||
func cartesianProduct(left, right [][]string, cap int) ([][]string, error) {
|
||||
if int64(len(left))*int64(len(right)) > int64(cap) {
|
||||
return nil, fmt.Errorf("DSL expansion exceeded %d combinations", cap)
|
||||
}
|
||||
result := make([][]string, 0, len(left)*len(right))
|
||||
for _, l := range left {
|
||||
for _, r := range right {
|
||||
combo := make([]string, 0, len(l)+len(r))
|
||||
combo = append(combo, l...)
|
||||
combo = append(combo, r...)
|
||||
result = append(result, combo)
|
||||
}
|
||||
}
|
||||
return result, nil
|
||||
}
|
||||
|
||||
// ParseAndExpandDSL tokenizes, parses, and expands a DSL string.
|
||||
// resolvedRefs contains previously expanded sets for +ref inlining.
|
||||
func ParseAndExpandDSL(dsl string, resolvedRefs map[string][][]string) ([][]string, error) {
|
||||
dsl = strings.TrimSpace(dsl)
|
||||
if dsl == "" {
|
||||
return nil, fmt.Errorf("empty DSL expression")
|
||||
}
|
||||
|
||||
tokens, err := tokenize(dsl)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("tokenize: %w", err)
|
||||
}
|
||||
|
||||
tree, err := parse(tokens)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("parse: %w", err)
|
||||
}
|
||||
|
||||
result, err := expand(tree, resolvedRefs)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
// Deduplicate models within each combination and sort for consistency
|
||||
for i, combo := range result {
|
||||
result[i] = dedupAndSort(combo)
|
||||
}
|
||||
|
||||
return result, nil
|
||||
}
|
||||
|
||||
// dedupAndSort removes duplicate entries and sorts alphabetically.
|
||||
func dedupAndSort(items []string) []string {
|
||||
seen := make(map[string]bool, len(items))
|
||||
var unique []string
|
||||
for _, item := range items {
|
||||
if !seen[item] {
|
||||
seen[item] = true
|
||||
unique = append(unique, item)
|
||||
}
|
||||
}
|
||||
sort.Strings(unique)
|
||||
return unique
|
||||
}
|
||||
|
||||
// extractRefs scans a DSL string for +ref tokens without full parsing.
|
||||
// Used for building the dependency graph for topological sorting.
|
||||
func extractRefs(dsl string) ([]string, error) {
|
||||
tokens, err := tokenize(dsl)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
var refs []string
|
||||
seen := make(map[string]bool)
|
||||
for _, t := range tokens {
|
||||
if t.typ == tokRef && !seen[t.val] {
|
||||
seen[t.val] = true
|
||||
refs = append(refs, t.val)
|
||||
}
|
||||
}
|
||||
return refs, nil
|
||||
}
|
||||
Reference in New Issue
Block a user