fix(vllm-cpp): resolve the DFlash draft path instead of missing the HF cache

The engine resolves speculative_config.model against a directory containing
config.json, or against ~/.cache/huggingface/hub/models--<org>--<repo>/
snapshots/*, and it never downloads. LocalAI keeps models in its own directory,
so the repo-id spelling the vLLM docs teach - "z-lab/Qwen3.6-27B-DFlash" - misses
the HF cache and dies deep inside the load with "draft checkpoint not found",
which reads like a broken checkpoint rather than a model nobody fetched.

Resolve it before the load call: the reference as given, then its last path
segment under LocalAI's models dir (what LocalAI's own downloader produces),
then the whole reference under the models dir. When none resolve, fail there
naming both what was asked for and every location tried, so the message says
what to do about it.

mtp and ngram pass through untouched - neither has a separate draft checkpoint.
A speculative_config that does not parse also passes through, because the engine
owns config validation and produces the better error.

Docs also gain the two limits that were missing and are easy to lose an
afternoon to: speculation is Qwen3.5/3.6-only at this engine pin regardless of
format, and mtp/dflash need a safetensors target. The latter is a gap in the
engine's GGUF loader rather than a property of GGUF - the format carries MTP
weights fine, llama.cpp reads them as nextn.* tensors plus a
<arch>.nextn_predict_layers key - so the docs say that rather than implying GGUF
cannot express it.

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

backend/go/vllm-cpp/backend.go

@@ -109,6 +109,16 @@ func (v *VllmCpp) Load(opts *pb.ModelOptions) error {
 
 	v.opts = parseOptions(opts)
 
+	// A DFlash draft is a second checkpoint the engine opens by path, and the
+	// engine never downloads one. Resolve it against LocalAI's models directory
+	// now so a repo-id spelling works, and so a missing draft fails here with an
+	// actionable message rather than as an HF-cache miss inside the load.
+	resolvedSpec, err := resolveDraftModelPath(v.opts.speculativeConfig, opts.ModelPath)
+	if err != nil {
+		return err
+	}
+	v.opts.speculativeConfig = resolvedSpec
+
 	mp := defaultModelParams()
 	if v.opts.blockSize > 0 {
 		mp.BlockSize = v.opts.blockSize

backend/go/vllm-cpp/options.go

@@ -18,6 +18,10 @@ package main
 
 import (
 	"encoding/json"
+	"fmt"
+	"os"
+	"path"
+	"path/filepath"
 	"strconv"
 	"strings"
 
@@ -206,6 +210,76 @@ func jsonInt32(v any, fallback int32) int32 {
 	}
 }
 
+// resolveDraftModelPath rewrites a DFlash draft reference into an absolute path
+// the engine can actually open.
+//
+// The engine resolves `speculative_config.model` against a directory containing
+// config.json, or against ~/.cache/huggingface/hub/models--<org>--<repo>/
+// snapshots/* - and it NEVER downloads. LocalAI keeps models in its own
+// directory, so a bare HF repo id (the spelling the vLLM docs teach) misses the
+// HF cache and dies deep in the load with "draft checkpoint not found", which
+// reads like a broken checkpoint rather than a missing download.
+//
+// So: try the reference as given, then the last path segment under the models
+// dir (`z-lab/Qwen3.6-27B-DFlash` -> `<models>/Qwen3.6-27B-DFlash`, which is
+// what LocalAI's own downloader produces), then the whole reference under the
+// models dir. If none exist, fail HERE with a message naming both what was
+// asked for and where we looked.
+//
+// mtp and ngram carry no separate draft checkpoint, so they pass through. A
+// document that does not parse also passes through: the engine owns config
+// validation and produces the better error.
+func resolveDraftModelPath(speculativeConfig, modelsDir string) (string, error) {
+	if strings.TrimSpace(speculativeConfig) == "" {
+		return speculativeConfig, nil
+	}
+	var spec map[string]any
+	if err := json.Unmarshal([]byte(speculativeConfig), &spec); err != nil {
+		return speculativeConfig, nil
+	}
+	if method, _ := spec["method"].(string); !strings.EqualFold(method, "dflash") {
+		return speculativeConfig, nil
+	}
+
+	ref, _ := spec["model"].(string)
+	ref = strings.TrimSpace(ref)
+	if ref == "" {
+		return "", fmt.Errorf(
+			"vllm-cpp: speculative_config method %q requires a \"model\" key naming the draft checkpoint", "dflash")
+	}
+
+	candidates := []string{ref}
+	if modelsDir != "" {
+		if base := path.Base(filepath.ToSlash(ref)); base != "" && base != "." && base != "/" {
+			candidates = append(candidates, filepath.Join(modelsDir, base))
+		}
+		candidates = append(candidates, filepath.Join(modelsDir, filepath.FromSlash(ref)))
+	}
+
+	for _, c := range candidates {
+		if _, err := os.Stat(filepath.Join(c, "config.json")); err != nil {
+			continue
+		}
+		abs, err := filepath.Abs(c)
+		if err != nil {
+			abs = c
+		}
+		spec["model"] = abs
+		out, err := json.Marshal(spec)
+		if err != nil {
+			return "", fmt.Errorf("vllm-cpp: re-encoding speculative_config: %w", err)
+		}
+		xlog.Info("[vllm-cpp] resolved DFlash draft checkpoint", "reference", ref, "path", abs)
+		return string(out), nil
+	}
+
+	return "", fmt.Errorf(
+		"vllm-cpp: DFlash draft checkpoint %q not found (looked in: %s). "+
+			"The engine does not download drafts - install the draft model into LocalAI first, "+
+			"or set speculative_config.model to an absolute path to a directory containing config.json",
+		ref, strings.Join(candidates, ", "))
+}
+
 func parseInt32(s string, fallback int32) int32 {
 	n, err := strconv.ParseInt(strings.TrimSpace(s), 10, 32)
 	if err != nil || n <= 0 {

backend/go/vllm-cpp/vllmcpp_test.go

@@ -265,6 +265,91 @@ var _ = Describe("samplingFromPredict", func() {
 	})
 })
 
+// The engine resolves speculative_config.model against a local directory or
+// ~/.cache/huggingface/hub ONLY - it never downloads. LocalAI keeps models in
+// its own directory, so a bare repo id would miss the HF cache and fail deep in
+// the load with a confusing "draft checkpoint not found". Resolve it here.
+var _ = Describe("resolveDraftModelPath", func() {
+	var modelsDir string
+
+	BeforeEach(func() {
+		modelsDir = GinkgoT().TempDir()
+	})
+
+	// draftDir creates a plausible draft checkpoint under models/.
+	draftDir := func(name string) string {
+		d := filepath.Join(modelsDir, name)
+		Expect(os.MkdirAll(d, 0o750)).To(Succeed())
+		Expect(os.WriteFile(filepath.Join(d, "config.json"), []byte("{}"), 0o600)).To(Succeed())
+		return d
+	}
+
+	It("rewrites a repo id to the matching directory in the models dir", func() {
+		want := draftDir("Qwen3.6-27B-DFlash")
+		spec := `{"method":"dflash","model":"z-lab/Qwen3.6-27B-DFlash"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(MatchJSON(`{"method":"dflash","model":"` + want + `"}`))
+	})
+
+	It("rewrites a models-dir-relative path", func() {
+		want := draftDir("drafts__dflash")
+		spec := `{"method":"dflash","model":"drafts__dflash"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(ContainSubstring(want))
+	})
+
+	It("leaves an absolute path that already resolves alone", func() {
+		abs := draftDir("elsewhere")
+		spec := `{"method":"dflash","model":"` + abs + `"}`
+		out, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(MatchJSON(spec))
+	})
+
+	It("fails with an actionable error when the draft is nowhere on disk", func() {
+		// Silently passing the repo id through would surface as an HF-cache
+		// miss inside the engine, which reads as "your model is broken".
+		spec := `{"method":"dflash","model":"z-lab/Not-Downloaded"}`
+		_, err := resolveDraftModelPath(spec, modelsDir)
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("z-lab/Not-Downloaded"))
+		Expect(err.Error()).To(ContainSubstring(modelsDir))
+	})
+
+	It("requires a model key for dflash", func() {
+		_, err := resolveDraftModelPath(`{"method":"dflash"}`, modelsDir)
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("model"))
+	})
+
+	It("leaves mtp and ngram configs untouched", func() {
+		// Neither has a separate draft checkpoint to resolve.
+		for _, spec := range []string{
+			`{"method":"mtp"}`,
+			`{"method":"ngram","num_speculative_tokens":4}`,
+		} {
+			out, err := resolveDraftModelPath(spec, modelsDir)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(out).To(MatchJSON(spec))
+		}
+	})
+
+	It("passes a malformed document through for the engine to reject", func() {
+		// The engine owns config validation and produces the better message.
+		out, err := resolveDraftModelPath(`{not json`, modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(Equal(`{not json`))
+	})
+
+	It("is a no-op on an empty config", func() {
+		out, err := resolveDraftModelPath("", modelsDir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(BeEmpty())
+	})
+})
+
 var _ = Describe("validModelPath", func() {
 	It("accepts a .gguf file", func() {
 		dir := GinkgoT().TempDir()

docs/content/features/text-generation.md

@@ -962,7 +962,22 @@ prefill from blowing up the per-step activation on the hybrid architectures.
 #### Speculative decoding
 
 `speculative_config:` takes the same JSON object as vLLM's
-`--speculative-config`. Three methods are supported:
+`--speculative-config`. Three methods are supported.
+
+> **Architecture limit.** At the current engine pin, `mtp` and `dflash` are
+> **Qwen3.5 / Qwen3.6 only**. The engine builds a widened speculative KV cache
+> directly for those families rather than through the model registry, so a
+> speculative config on any other architecture (Llama, GLM, Gemma, Mistral, ...)
+> will not work regardless of checkpoint format. `ngram` needs no draft weights
+> and is not subject to this limit.
+
+> **Format limit.** `mtp` and `dflash` require a **safetensors** target and are
+> rejected at load on a `.gguf` target, with:
+> `speculative decoding requires a safetensors target checkpoint`.
+> This is a current gap in the engine's GGUF loader, not a property of the GGUF
+> format - GGUF can carry MTP weights (llama.cpp reads them as `nextn.*`
+> tensors plus a `<arch>.nextn_predict_layers` key), but vllm.cpp's GGUF path
+> does not map them yet. `ngram` works fine on GGUF.
 
 **MTP** (Multi-Token Prediction) uses a draft head shipped inside the target
 checkpoint's own `mtp.*` tensors, so there is no second model to download. It
@@ -990,6 +1005,18 @@ engine_args:
     num_speculative_tokens: 4
 ```
 
+The draft shares the *target's* `embed_tokens` and `lm_head`, so both must come
+from the same model family and the target must be safetensors.
+
+**The engine does not download the draft.** `model:` is resolved, in order,
+as a path as given, then as the last path segment under LocalAI's models
+directory (`z-lab/Qwen3.6-27B-DFlash` → `<models>/Qwen3.6-27B-DFlash`, which is
+what LocalAI's own downloader produces), then as the whole reference under the
+models directory. Install the draft into LocalAI first, or give an absolute path
+to a directory containing `config.json`. If none of those resolve, the load
+fails immediately naming every location that was tried, rather than reporting a
+missing checkpoint from inside the engine.
+
 **N-gram** needs no draft model at all - it proposes from the prompt's own
 suffix history. `num_speculative_tokens` is required: