feat(vllm-cpp): wire the full engine config surface through engine_args

The vllm-cpp backend could configure four of the engine's knobs - block size,
KV block count, max sequence length and max concurrent sequences - out of a
config surface that is considerably larger. Speculative decoding, prefix
caching, the chunked-prefill token budget, the scheduling policy and the
external KV connector were all reachable from vllm.cpp's own HTTP server and
from nothing LocalAI could write in a model config.

Part of that gap was the C ABI itself, which carried strictly less than
EngineParams does; that is fixed upstream in vllm.cpp ABI v9 (this bumps the pin
to it). The rest was here: the backend parsed a flat `options:` list with five
recognised keys and had no way to express a nested JSON document at all.

Configuration now goes through `engine_args:`, the same map the vLLM and SGLang
backends already take, with keys spelled as vLLM's own CLI flags - so a
`speculative_config` or `kv_transfer_config` block written for vLLM works
verbatim:

  engine_args:
    max_num_batched_tokens: 8192
    enable_prefix_caching: true
    scheduling_policy: lpm
    speculative_config:
      method: dflash
      model: z-lab/Qwen3.6-27B-DFlash
      num_speculative_tokens: 4
    kv_transfer_config:
      kv_connector: LMCacheConnector
      kv_role: kv_both
      kv_connector_extra_config: {host: 127.0.0.1, port: 65432}

The `options:` list keeps working, and now reads every key too, so no existing
config breaks; engine_args wins where both set the same key.

Two details worth calling out. `enable_prefix_caching: false` maps to the ABI's
force-OFF state (2), not the 0 that means "let the model capability decide" -
collapsing them would silently turn the cache ON for the dense architectures
that default it on. And cSamplingParams grows the ABI v8 logits-processor tail:
LocalAI installs no processor, but the C side reads those fields off the pointer
we hand it, so a Go struct that stopped short would have had the engine read 16
bytes past our allocation and call whatever sat there.

The importer gets the safetensors counterpart of the llama-cpp MTP hook: a
`vllm-cpp` import of a HuggingFace repo probes config.json and, on a checkpoint
that declares an MTP head, writes speculative_config {method: mtp} into the
generated engine_args. DFlash draft repositories are detected and refused with a
warning rather than configured as standalone models, since a drafter cannot
serve alone. The llama-cpp importer stops applying its own `spec_type:draft-mtp`
options when the chosen backend is vllm-cpp: those are llama.cpp option keys
vllm-cpp does not read, and vllm.cpp rejects MTP over a GGUF source anyway
because the `mtp.*` draft tensors only exist in the safetensors checkpoint.

docs/content/features/text-generation.md gains a vllm.cpp section - the backend
had no documentation page at all - covering the engine_args table, all three
speculative methods, LMCache, and the legacy options list.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-opus-5 [ClaudeCode]

Author: mudler

backend/go/vllm-cpp/Makefile

@@ -11,7 +11,7 @@ JOBS?=$(shell nproc --ignore=1 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || e
 
 # vllm.cpp version
 VLLM_CPP_REPO?=https://github.com/mudler/vllm.cpp
-VLLM_CPP_VERSION?=9e1c9025ae61167a3335454d7cc0de6093c21845
+VLLM_CPP_VERSION?=974d9d72c2365d20dc76dbce4a98fddac32525c0
 
 # The backend consumes only the stable C ABI (libvllm + include/vllm.h), so the
 # server, examples and tests of the engine are never built here.

backend/go/vllm-cpp/backend.go

@@ -116,34 +116,60 @@ func (v *VllmCpp) Load(opts *pb.ModelOptions) error {
 	if v.opts.numBlocks > 0 {
 		mp.NumBlocks = v.opts.numBlocks
 	}
+	// Sequence-length precedence, narrowest source last: context_size is the
+	// generic LocalAI knob every backend honours, max_model_len is the
+	// vLLM-specific one, and engine_args.max_model_len is the explicit
+	// vllm-cpp override.
 	if opts.ContextSize > 0 {
 		mp.MaxModelLen = opts.ContextSize
 	}
+	if opts.MaxModelLen > 0 {
+		mp.MaxModelLen = opts.MaxModelLen
+	}
+	if v.opts.maxModelLen > 0 {
+		mp.MaxModelLen = v.opts.maxModelLen
+	}
 	if v.opts.maxNumSeqs > 0 {
 		mp.MaxNumSeqs = v.opts.maxNumSeqs
 	}
+	if v.opts.maxNumBatchedTokens > 0 {
+		mp.MaxNumBatchedTokens = v.opts.maxNumBatchedTokens
+	}
+	mp.EnablePrefixCaching = v.opts.enablePrefixCaching
 
+	// Every string below is borrowed by C for the duration of the load call
+	// only (the library copies what it keeps), so the backing slices just have
+	// to outlive vllmEngineLoad - hence the single KeepAlive after it.
 	modelC := cString(model)
 	mp.ModelPath = uintptr(unsafe.Pointer(&modelC[0])) // #nosec G103 -- borrowed by C for the load call only
-	var toolParserC, reasoningParserC []byte
-	if v.opts.toolParser != "" {
-		toolParserC = cString(v.opts.toolParser)
-		mp.ToolParser = uintptr(unsafe.Pointer(&toolParserC[0])) // #nosec G103 -- borrowed by C for the load call only
-	}
-	if v.opts.reasoningParser != "" {
-		reasoningParserC = cString(v.opts.reasoningParser)
-		mp.ReasoningParser = uintptr(unsafe.Pointer(&reasoningParserC[0])) // #nosec G103 -- borrowed by C for the load call only
+	keep := [][]byte{modelC}
+	setStr := func(dst *uintptr, s string) {
+		if s == "" {
+			return
+		}
+		b := cString(s)
+		keep = append(keep, b)
+		*dst = uintptr(unsafe.Pointer(&b[0])) // #nosec G103 -- borrowed by C for the load call only
 	}
+	setStr(&mp.ToolParser, v.opts.toolParser)
+	setStr(&mp.ReasoningParser, v.opts.reasoningParser)
+	setStr(&mp.SpeculativeConfig, v.opts.speculativeConfig)
+	setStr(&mp.KVTransferConfig, v.opts.kvTransferConfig)
+	setStr(&mp.SchedulingPolicy, v.opts.schedulingPolicy)
+	setStr(&mp.TokenizerConfigPath, v.opts.tokenizerConfigPath)
 
 	xlog.Info("[vllm-cpp] Load", "model", model, "engine", vllmVersion(),
 		"blockSize", mp.BlockSize, "numBlocks", mp.NumBlocks,
-		"maxModelLen", mp.MaxModelLen, "maxNumSeqs", mp.MaxNumSeqs)
+		"maxModelLen", mp.MaxModelLen, "maxNumSeqs", mp.MaxNumSeqs,
+		"maxNumBatchedTokens", mp.MaxNumBatchedTokens,
+		"prefixCaching", prefixCachingName(mp.EnablePrefixCaching),
+		"schedulingPolicy", v.opts.schedulingPolicy,
+		"speculativeConfig", v.opts.speculativeConfig,
+		"kvTransferConfig", v.opts.kvTransferConfig)
 
 	var engine uintptr
 	rc := vllmEngineLoad(unsafe.Pointer(&mp), unsafe.Pointer(&engine)) // #nosec G103 -- POD out-params
-	runtime.KeepAlive(modelC)
-	runtime.KeepAlive(toolParserC)
-	runtime.KeepAlive(reasoningParserC)
+	runtime.KeepAlive(keep)
 	if rc != vllmOK {
 		return fmt.Errorf("vllm-cpp: engine load failed: %s", vllmLastError())
 	}

backend/go/vllm-cpp/govllmcpp.go

@@ -1,6 +1,6 @@
 package main
 
-// purego bindings for the vllm.cpp stable C ABI (include/vllm.h, ABI v2).
+// purego bindings for the vllm.cpp stable C ABI (include/vllm.h, ABI v9).
 //
 // The structs below are hand-mirrored PODs of the C declarations, with
 // explicit padding so the Go layout matches the C layout on linux/darwin
@@ -18,23 +18,52 @@ import (
 )
 
 // abiVersion is the VLLM_ABI_VERSION this file mirrors (vllm.h).
-const abiVersion = 5
+const abiVersion = 9
+
+// Prefix-caching tri-state (vllm_model_params.enable_prefix_caching, ABI v7).
+// 0 is the model-capability default, which is what a config that says nothing
+// about prefix caching must produce.
+const (
+	prefixCachingModelDefault int32 = 0
+	prefixCachingOn           int32 = 1
+	prefixCachingOff          int32 = 2
+)
+
+// prefixCachingName renders the tri-state for the load log line, where "0"
+// would otherwise read as "off" rather than "whatever the model defaults to".
+func prefixCachingName(state int32) string {
+	switch state {
+	case prefixCachingOn:
+		return "on"
+	case prefixCachingOff:
+		return "off"
+	default:
+		return "model-default"
+	}
+}
 
 // vllm_status (vllm.h).
 const (
 	vllmOK = 0
 )
 
-// cModelParams mirrors vllm_model_params.
+// cModelParams mirrors vllm_model_params. Every field is naturally aligned on
+// LP64 (the int32 pairs sit together), so the Go layout needs no explicit
+// padding; the offsets are asserted in vllmcpp_test.go.
 type cModelParams struct {
 	ModelPath           uintptr // const char*
-	TokenizerConfigPath uintptr // const char*
+	TokenizerConfigPath uintptr // const char*; NULL = <model_dir>/... (ABI v9)
 	BlockSize           int32
 	NumBlocks           int32
 	MaxModelLen         int32
 	MaxNumSeqs          int32
 	ToolParser          uintptr // const char*; NULL = auto-detect (ABI v4)
 	ReasoningParser     uintptr // const char*; NULL = auto-detect (ABI v5)
+	SpeculativeConfig   uintptr // const char* JSON; NULL = no speculation (ABI v6)
+	EnablePrefixCaching int32   // tri-state 0/1/2 (ABI v7)
+	MaxNumBatchedTokens int32   // <= 0 = per-arch default (ABI v9)
+	SchedulingPolicy    uintptr // const char*; NULL = "fcfs" (ABI v9)
+	KVTransferConfig    uintptr // const char* JSON; NULL = no connector (ABI v9)
 }
 
 // cSamplingParams mirrors vllm_sampling_params (ABI v2, structured fields
@@ -65,6 +94,12 @@ type cSamplingParams struct {
 	StructuredGrammar    uintptr // const char*
 	StructuredJSONObject int32
 	_                    [4]byte
+	// ABI v8 tail. LocalAI installs no custom logits processor, but the fields
+	// MUST be mirrored: the C side reads them off the pointer we hand it, so a
+	// Go struct that stopped at StructuredJSONObject would have the engine read
+	// 16 bytes past our allocation and call whatever garbage sat there.
+	LogitsProcessor         uintptr // vllm_logits_processor; NULL = none
+	LogitsProcessorUserData uintptr // void*
 }
 
 // cCompletion mirrors vllm_completion.

backend/go/vllm-cpp/options.go

@@ -1,30 +1,72 @@
 package main
 
-// Engine-sizing knobs carried through the model config's free-form
-// `options:` list ("key:value" entries), mirroring how the other in-house
-// backends pass engine-specific settings that have no proto field.
+// Load-time engine configuration, from two config surfaces:
+//
+//   - `engine_args:` (ModelOptions.EngineArgs, a JSON object) is the canonical
+//     one. Keys are spelled exactly as vLLM's own CLI flags, so a config written
+//     against vLLM works verbatim here - `speculative_config` and
+//     `kv_transfer_config` in particular take the same JSON documents vLLM's
+//     --speculative-config / --kv-transfer-config accept, and are handed to the
+//     engine unparsed.
+//   - `options:` (the free-form "key:value" list) is the older surface this
+//     backend shipped with. It is still honoured so existing configs keep
+//     working; engine_args wins on any key set in both.
+//
+// Anything unrecognised is ignored rather than fatal: the engine validates the
+// documents it is given and reports a precise error at load, and a config that
+// also carries knobs for a different backend must not fail the load here.
 
 import (
+	"encoding/json"
 	"strconv"
 	"strings"
 
 	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/xlog"
 )
 
 type loadOptions struct {
 	blockSize  int32 // KV block size (tokens/block); engine default 32.
 	numBlocks  int32 // KV blocks to allocate; engine default 256.
 	maxNumSeqs int32 // max concurrent sequences; engine default 8.
+	// Max sequence length. Also settable through the model config's
+	// context_size / max_model_len; see Load for the precedence.
+	maxModelLen int32
+	// Per-step chunked-prefill token budget (ABI v9). 0 = the engine's
+	// bounded per-arch default.
+	maxNumBatchedTokens int32
+	// Automatic prefix caching tri-state (ABI v7): 0 = the model-capability
+	// default, 1 = force on, 2 = force off.
+	enablePrefixCaching int32
+	// Scheduler admission policy (ABI v9): "" = fcfs, else fcfs|priority|lpm.
+	schedulingPolicy string
 	// Engine-side parser selection (ABI v4/v5). Empty = the engine
 	// auto-detects from the chat template; "none" disables the reasoning
 	// split; unknown names fail the first chat call.
 	toolParser      string
 	reasoningParser string
+	// Speculative decoding (ABI v6), as vLLM's --speculative-config JSON:
+	// {"method":"mtp"|"dflash"|"ngram", ...}. Empty = no speculation.
+	speculativeConfig string
+	// External KV connector / LMCache (ABI v9), as vLLM's --kv-transfer-config
+	// JSON. Empty = no connector.
+	kvTransferConfig string
+	// Override for the tokenizer_config.json the chat template is read from
+	// (ABI v9). Empty = <model_dir>/tokenizer_config.json.
+	tokenizerConfigPath string
 }
 
 func parseOptions(opts *pb.ModelOptions) loadOptions {
 	lo := loadOptions{}
-	for _, o := range opts.GetOptions() {
+	applyOptionsList(&lo, opts.GetOptions())
+	applyEngineArgs(&lo, opts.GetEngineArgs())
+	return lo
+}
+
+// applyOptionsList reads the legacy free-form "key:value" list. strings.Cut
+// splits on the FIRST colon only, so a JSON object value survives intact.
+func applyOptionsList(lo *loadOptions, options []string) {
+	for _, o := range options {
 		k, v, found := strings.Cut(o, ":")
 		if !found {
 			continue
@@ -36,13 +78,132 @@ func parseOptions(opts *pb.ModelOptions) loadOptions {
 			lo.numBlocks = parseInt32(v, lo.numBlocks)
 		case "max_num_seqs":
 			lo.maxNumSeqs = parseInt32(v, lo.maxNumSeqs)
-		case "tool_parser":
+		case "max_num_batched_tokens":
+			lo.maxNumBatchedTokens = parseInt32(v, lo.maxNumBatchedTokens)
+		case "max_model_len":
+			lo.maxModelLen = parseInt32(v, lo.maxModelLen)
+		case "scheduling_policy", "schedule_policy":
+			lo.schedulingPolicy = strings.TrimSpace(v)
+		case "tool_parser", "tool_call_parser":
 			lo.toolParser = strings.TrimSpace(v)
 		case "reasoning_parser":
 			lo.reasoningParser = strings.TrimSpace(v)
+		case "speculative_config":
+			lo.speculativeConfig = strings.TrimSpace(v)
+		case "kv_transfer_config":
+			lo.kvTransferConfig = strings.TrimSpace(v)
+		case "tokenizer_config", "tokenizer_config_path":
+			lo.tokenizerConfigPath = strings.TrimSpace(v)
+		case "enable_prefix_caching", "enable_radix_attention":
+			if b, err := strconv.ParseBool(strings.TrimSpace(v)); err == nil {
+				lo.enablePrefixCaching = prefixCachingTriState(b)
+			}
 		}
 	}
-	return lo
+}
+
+// applyEngineArgs overlays the `engine_args:` JSON object. A document that does
+// not parse is logged and skipped: engine_args is shared with the other engines
+// (the vLLM and SGLang backends read the same field), so a stray key must not
+// take the model down.
+func applyEngineArgs(lo *loadOptions, engineArgs string) {
+	if strings.TrimSpace(engineArgs) == "" {
+		return
+	}
+	var args map[string]any
+	if err := json.Unmarshal([]byte(engineArgs), &args); err != nil {
+		xlog.Warn("[vllm-cpp] ignoring unparseable engine_args", "error", err)
+		return
+	}
+	for k, v := range args {
+		switch k {
+		case "block_size":
+			lo.blockSize = jsonInt32(v, lo.blockSize)
+		case "num_blocks":
+			lo.numBlocks = jsonInt32(v, lo.numBlocks)
+		case "max_num_seqs":
+			lo.maxNumSeqs = jsonInt32(v, lo.maxNumSeqs)
+		case "max_num_batched_tokens":
+			lo.maxNumBatchedTokens = jsonInt32(v, lo.maxNumBatchedTokens)
+		case "max_model_len":
+			lo.maxModelLen = jsonInt32(v, lo.maxModelLen)
+		case "scheduling_policy", "schedule_policy":
+			lo.schedulingPolicy = jsonString(v, lo.schedulingPolicy)
+		case "tool_parser", "tool_call_parser":
+			lo.toolParser = jsonString(v, lo.toolParser)
+		case "reasoning_parser":
+			lo.reasoningParser = jsonString(v, lo.reasoningParser)
+		case "tokenizer_config", "tokenizer_config_path":
+			lo.tokenizerConfigPath = jsonString(v, lo.tokenizerConfigPath)
+		case "speculative_config":
+			lo.speculativeConfig = jsonDocument(v, lo.speculativeConfig, k)
+		case "kv_transfer_config":
+			lo.kvTransferConfig = jsonDocument(v, lo.kvTransferConfig, k)
+		case "enable_prefix_caching", "enable_radix_attention":
+			if b, ok := v.(bool); ok {
+				lo.enablePrefixCaching = prefixCachingTriState(b)
+			}
+		default:
+			xlog.Debug("[vllm-cpp] ignoring unknown engine_args key", "key", k)
+		}
+	}
+}
+
+// prefixCachingTriState maps a YAML/JSON boolean onto the C ABI's tri-state. An
+// explicit `false` must reach the engine as force-OFF (2), NOT as the 0 that
+// means "let the model capability decide" - those differ for the hybrid archs,
+// which default the cache off, and for the dense ones, which default it on.
+func prefixCachingTriState(on bool) int32 {
+	if on {
+		return prefixCachingOn
+	}
+	return prefixCachingOff
+}
+
+// jsonDocument normalises an object-valued engine_args entry to a JSON string
+// for the C ABI. YAML nesting arrives as a map (the natural spelling); a
+// pre-encoded JSON string is accepted too, since a config round-tripped through
+// a flat store may carry it that way.
+func jsonDocument(v any, fallback string, key string) string {
+	switch t := v.(type) {
+	case string:
+		if strings.TrimSpace(t) == "" {
+			return fallback
+		}
+		return t
+	default:
+		buf, err := json.Marshal(t)
+		if err != nil {
+			xlog.Warn("[vllm-cpp] ignoring unencodable engine_args value", "key", key, "error", err)
+			return fallback
+		}
+		return string(buf)
+	}
+}
+
+func jsonString(v any, fallback string) string {
+	s, ok := v.(string)
+	if !ok {
+		return fallback
+	}
+	return strings.TrimSpace(s)
+}
+
+// jsonInt32 accepts the float64 a JSON number decodes to, plus the string
+// spelling a YAML config may produce. Non-positive values keep the fallback:
+// every knob this covers uses "<= 0 means the engine default".
+func jsonInt32(v any, fallback int32) int32 {
+	switch t := v.(type) {
+	case float64:
+		if t <= 0 || t > 1<<31-1 {
+			return fallback
+		}
+		return int32(t)
+	case string:
+		return parseInt32(t, fallback)
+	default:
+		return fallback
+	}
 }
 
 func parseInt32(s string, fallback int32) int32 {