hf: use the reasoning parser a card's SGLang command names

CHANGELOG.md

@@ -9,6 +9,17 @@ the fact.
 
 ## [Unreleased]
 
+### Fixed
+
+- Profiles for Qwen3 hybrid models got `--reasoning-parser deepseek_r1` from
+  the card's `vllm serve` line. With thinking turned off
+  (`enable_thinking: false`), a non-streaming reply then came back with
+  `content` empty and the answer in `reasoning`. When the card's SGLang
+  command names a different parser and the installed vLLM has one by that
+  name (Qwen3's cards name `qwen3`), that parser is used, and `inspect` and
+  `new` say so. Existing profiles keep their parser; change one with
+  `llm-helper set`.
+
 ## [0.6.0] - 2026-11-04
 
 ### Added

docs/models.md

@@ -12,6 +12,8 @@ llm-helper reads a model's HuggingFace card before anything is saved: it checks
 - `llm-helper inspect <org/model>` shows one repo's license, its weight size against this machine's budget, and the vLLM flags its card suggests — without pulling anything.
 
 `search`, `inspect`, and `new` pre-fill known vLLM serving flags from the model card: tool call parsers, reasoning parsers, and context length.
+When the card's SGLang command names a different reasoning parser than its `vllm serve` line, and the installed vLLM has a parser by that exact name, that one is used and `inspect` says so.
+Qwen3's cards are the case this covers: their `vllm serve` line names `deepseek_r1`, which puts the whole answer in `reasoning` when a request turns thinking off.
 `search` marks what is too big for this machine and what cannot load, and preselects the biggest quant that fits.
 `new` refuses a quant that cannot load rather than serving it — activating a profile restarts the service, so a model that cannot load would take the running one down with it.
 

internal/catalog/catalog.go

@@ -102,6 +102,7 @@ type Detail struct {
 	// vLLM won't take, so it's visible before pulling rather than as a
 	// crash loop afterwards.
 	DroppedFlags []string `json:"dropped_flags,omitempty"`
+	FlagNotes    []string `json:"flag_notes,omitempty"`
 	// KVBytesPerToken is what one token of context costs in KV cache, and
 	// MaxContext the longest the model was trained for. Both 0 when the
 	// config could not be read.
@@ -230,12 +231,27 @@ func Inspect(ctx context.Context, m hf.Model, budget detect.WeightsBudget, caps
 	if flags, err := hf.DetectFlags(ctx, m.ID); err == nil {
 		// Validate here too: seeing that a card's parser won't work on this
 		// machine is most useful before pulling tens of gigabytes.
-		d.DetectedFlags, d.DroppedFlags = vllm.Check(flags.ExtraArgs, caps)
+		args, note := cardFlags(flags, caps)
+		if note != "" {
+			d.FlagNotes = append(d.FlagNotes, note)
+		}
+		d.DetectedFlags, d.DroppedFlags = vllm.Check(args, caps)
 		d.MaxModelLen = flags.MaxModelLen
 	}
 	return d
 }
 
+// cardFlags returns the card's vllm serve flags, rebuilt around the parser its
+// SGLang command names when this vLLM registers that name.
+func cardFlags(flags hf.DetectedFlags, caps vllm.Caps) (args, note string) {
+	if flags.SGLangExtraArgs == "" || !caps.Known() || !caps.HasReasoningParser(flags.SGLangReasoningParser) {
+		return flags.ExtraArgs, ""
+	}
+	return flags.SGLangExtraArgs, fmt.Sprintf(
+		"using --reasoning-parser %s instead of the card's %s: the card's SGLang command names %s, and vLLM %s has it",
+		flags.SGLangReasoningParser, flags.ReasoningParser, flags.SGLangReasoningParser, caps.Version)
+}
+
 // ForProvider reshapes what Inspect found for the server that will load the
 // model, since Inspect judges GGUF by vLLM's plugin. A server that picks its
 // own GGUF quant has none of the plugin's limits, so split and subfolder
@@ -255,6 +271,7 @@ func ForProvider(d Detail, prov provider.Provider) Detail {
 	d.ArchBlocked = ""
 	d.DetectedFlags = ""
 	d.DroppedFlags = nil
+	d.FlagNotes = nil
 	if !caps.GGUFQuant && d.NeedsGGUF {
 		d.ArchBlocked = "only published as GGUF here, which " + prov.Name() + " does not load"
 	}
@@ -357,7 +374,11 @@ func PrepareProfile(ctx context.Context, dirs config.Dirs, repo string, choice *
 	// author prescribed, the other is what this tool added. Merging them
 	// before reporting attributed our own --tokenizer to the card, and said
 	// "flags auto-detected from README" about a repo whose README had none.
-	cardArgs, dropped := vllm.Check(detected.ExtraArgs, caps)
+	cardArgs, note := cardFlags(detected, caps)
+	if note != "" {
+		p.Notes = append(p.Notes, note)
+	}
+	cardArgs, dropped := vllm.Check(cardArgs, caps)
 	p.Notes = append(p.Notes, dropped...)
 
 	if patched, needsPatch, err := hf.DetectChatTemplatePatch(ctx, repo); err != nil {

internal/catalog/catalog_internal_test.go

@@ -763,3 +763,40 @@ func TestAPullWithoutAQuantChoiceStillMeasuresTheWeights(t *testing.T) {
 		t.Error("the pull planned a KV cache without ever measuring the model it has to fit around")
 	}
 }
+
+func TestCardFlagsPrefersARegisteredSGLangParser(t *testing.T) {
+	t.Parallel()
+	qwen := hf.DetectedFlags{
+		ExtraArgs:             "--reasoning-parser deepseek_r1",
+		ReasoningParser:       "deepseek_r1",
+		SGLangReasoningParser: "qwen3",
+		SGLangExtraArgs:       "--reasoning-parser qwen3",
+	}
+	unregistered := qwen
+	unregistered.SGLangReasoningParser = "deepseek-r1"
+	unregistered.SGLangExtraArgs = "--reasoning-parser deepseek-r1"
+	tests := []struct {
+		name    string
+		flags   hf.DetectedFlags
+		caps    vllm.Caps
+		want    string
+		changed bool
+	}{
+		{"registered", qwen, spark, "--reasoning-parser qwen3", true},
+		{"not registered", unregistered, spark, "--reasoning-parser deepseek_r1", false},
+		{"caps unknown", qwen, vllm.Caps{}, "--reasoning-parser deepseek_r1", false},
+		{"parsers agree", hf.DetectedFlags{ExtraArgs: "--reasoning-parser qwen3", SGLangReasoningParser: "qwen3"}, spark, "--reasoning-parser qwen3", false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			got, note := cardFlags(tt.flags, tt.caps)
+			if got != tt.want {
+				t.Errorf("args = %q, want %q", got, tt.want)
+			}
+			if tt.changed != (note != "") {
+				t.Errorf("note = %q, want changed=%v", note, tt.changed)
+			}
+		})
+	}
+}

internal/cli/search.go

@@ -260,6 +260,9 @@ func printModelDetail(d catalog.Detail, cfg *config.Config) {
 	} else {
 		fmt.Println("  vllm flags: no vllm serve snippet in README -- set flags by hand after pulling")
 	}
+	for _, n := range d.FlagNotes {
+		fmt.Printf("  changed:    %s\n", n)
+	}
 	for _, w := range d.DroppedFlags {
 		fmt.Printf("  dropped:    %s\n", w)
 	}

internal/hf/hf.go

@@ -690,6 +690,14 @@ func nextPageURL(header, want string) string {
 type DetectedFlags struct {
 	ExtraArgs   string // e.g. "--enable-auto-tool-choice --tool-call-parser qwen3_coder"
 	MaxModelLen string // from --max-model-len in the snippet, "" if absent
+	// SGLangReasoningParser is the --reasoning-parser the card's SGLang
+	// command names, "" if none or if its SGLang commands disagree.
+	SGLangReasoningParser string
+	// ReasoningParser is the --reasoning-parser the vllm serve line names.
+	ReasoningParser string
+	// SGLangExtraArgs is ExtraArgs rebuilt with SGLangReasoningParser, set
+	// only when the two parsers differ.
+	SGLangExtraArgs string
 }
 
 // DetectFlags fetches a model's README and pulls known vLLM serving flags
@@ -717,8 +725,12 @@ func DetectFlags(ctx context.Context, repo string) (DetectedFlags, error) {
 	// context clamp below -- those models are exactly the ones whose real
 	// context is far under the generic default.
 	snippetRe := regexp.MustCompile(`vllm serve(?:.*\\\n)*.*`)
+	d.SGLangReasoningParser = sglangReasoningParser(string(body))
 	if snippet := snippetRe.FindString(string(body)); snippet != "" {
-		d.ExtraArgs, d.MaxModelLen = flagsFromSnippet(snippet)
+		d.ExtraArgs, d.MaxModelLen, d.ReasoningParser = parseSnippet(snippet, "")
+		if d.ReasoningParser != "" && d.SGLangReasoningParser != "" && d.ReasoningParser != d.SGLangReasoningParser {
+			d.SGLangExtraArgs, _, _ = parseSnippet(snippet, d.SGLangReasoningParser)
+		}
 	}
 
 	// Model cards routinely paste a snippet from a bigger sibling model, so

internal/hf/hf_internal_test.go

@@ -1138,3 +1138,25 @@ func TestGGUFServable(t *testing.T) {
 		t.Error("an unreadable config was treated as a refusal")
 	}
 }
+
+func TestDetectFlagsReadsTheSGLangParser(t *testing.T) {
+	serveTestdata(t, map[string]string{
+		"/" + repo + "/raw/main/README.md": "README_qwen3_hybrid.md",
+	})
+	d, err := DetectFlags(context.Background(), repo)
+	if err != nil {
+		t.Fatalf("DetectFlags: %v", err)
+	}
+	if d.ExtraArgs != "--enable-reasoning --reasoning-parser deepseek_r1" {
+		t.Errorf("ExtraArgs = %q", d.ExtraArgs)
+	}
+	if d.SGLangReasoningParser != "qwen3" {
+		t.Errorf("SGLangReasoningParser = %q, want qwen3", d.SGLangReasoningParser)
+	}
+	if d.ReasoningParser != "deepseek_r1" {
+		t.Errorf("ReasoningParser = %q, want deepseek_r1", d.ReasoningParser)
+	}
+	if d.SGLangExtraArgs != "--enable-reasoning --reasoning-parser qwen3" {
+		t.Errorf("SGLangExtraArgs = %q", d.SGLangExtraArgs)
+	}
+}

internal/hf/snippet.go

@@ -53,6 +53,13 @@ var thinkingParsers = regexp.MustCompile(`^(deepseek_v3|poolside_v1)$`)
 // name is real depends on the vLLM installed on this machine, which is not
 // something a HuggingFace client should know about.
 func flagsFromSnippet(snippet string) (extraArgs, maxModelLen string) {
+	extraArgs, maxModelLen, _ = parseSnippet(snippet, "")
+	return extraArgs, maxModelLen
+}
+
+// parseSnippet is flagsFromSnippet that also returns the --reasoning-parser it
+// kept, replaced by parser when that is set.
+func parseSnippet(snippet, parser string) (extraArgs, maxModelLen, reasoningParser string) {
 	// Fold shell line continuations, then flatten what is left. The backslash
 	// and the newline are separate concerns: a card can wrap without one, and
 	// a bare newline left in place gets read as the previous flag's value.
@@ -62,7 +69,7 @@ func flagsFromSnippet(snippet string) (extraArgs, maxModelLen string) {
 	if err != nil {
 		// A card whose snippet isn't valid shell (prose wrapped mid-line,
 		// an unbalanced quote) yields nothing rather than half of it.
-		return "", ""
+		return "", "", ""
 	}
 
 	var kept []string
@@ -92,6 +99,9 @@ func flagsFromSnippet(snippet string) (extraArgs, maxModelLen string) {
 			continue
 		}
 		if flag == "--reasoning-parser" {
+			if parser != "" {
+				value = parser
+			}
 			sawParser = value
 		}
 		if value == "" {
@@ -104,7 +114,42 @@ func flagsFromSnippet(snippet string) (extraArgs, maxModelLen string) {
 	if thinkingParsers.MatchString(sawParser) {
 		kept = append(kept, `--default-chat-template-kwargs {"thinking":true}`)
 	}
-	return strings.Join(kept, " "), maxModelLen
+	return strings.Join(kept, " "), maxModelLen, sawParser
+}
+
+var sglangSnippetRe = regexp.MustCompile(`sglang[._](?:launch_server|serve)(?:.*\\\n)*.*`)
+
+func sglangReasoningParser(card string) string {
+	found := ""
+	for _, snippet := range sglangSnippetRe.FindAllString(card, -1) {
+		name := reasoningParserIn(snippet)
+		if name == "" {
+			continue
+		}
+		if found != "" && found != name {
+			return ""
+		}
+		found = name
+	}
+	return found
+}
+
+func reasoningParserIn(snippet string) string {
+	flat := strings.ReplaceAll(snippet, "\\\n", " ")
+	flat = strings.NewReplacer("\r", " ", "\n", " ", "\t", " ").Replace(flat)
+	args, err := config.SplitArgs(flat)
+	if err != nil {
+		return ""
+	}
+	for i, a := range args {
+		if v, ok := strings.CutPrefix(a, "--reasoning-parser="); ok {
+			return v
+		}
+		if a == "--reasoning-parser" && i+1 < len(args) && !isFlagToken(args[i+1]) {
+			return args[i+1]
+		}
+	}
+	return ""
 }
 
 // looksSecret keeps a credential pasted into a card out of a profile file.

internal/hf/snippet_internal_test.go

@@ -93,3 +93,45 @@ func TestFlagsFromSnippetUnparseable(t *testing.T) {
 		t.Errorf("got %q, want nothing from an unparseable snippet", got)
 	}
 }
+
+func TestParseSnippetReplacesTheParser(t *testing.T) {
+	t.Parallel()
+	tests := []struct {
+		name, snippet, parser, want string
+	}{
+		{"into deepseek_v3", `vllm serve org/m --reasoning-parser deepseek_r1`, "deepseek_v3", `--reasoning-parser deepseek_v3 --default-chat-template-kwargs {"thinking":true}`},
+		{"away from deepseek_v3", `vllm serve org/m --reasoning-parser deepseek_v3`, "qwen3", "--reasoning-parser qwen3"},
+		{"equals form", `vllm serve org/m --reasoning-parser=deepseek_r1 --enforce-eager`, "qwen3", "--reasoning-parser qwen3 --enforce-eager"},
+		{"no parser in the snippet", `vllm serve org/m --enforce-eager`, "qwen3", "--enforce-eager"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			got, _, _ := parseSnippet(tt.snippet, tt.parser)
+			if got != tt.want {
+				t.Errorf("parseSnippet() = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestSGLangReasoningParser(t *testing.T) {
+	t.Parallel()
+	tests := []struct {
+		name, card, want string
+	}{
+		{"qwen3 hybrid card", "- SGLang:\n    ```shell\n    python -m sglang.launch_server --model-path Qwen/Qwen3-0.6B --reasoning-parser qwen3\n    ```\n- vLLM:\n    ```shell\n    vllm serve Qwen/Qwen3-0.6B --enable-reasoning --reasoning-parser deepseek_r1\n    ```\n", "qwen3"},
+		{"continued lines", "python3 -m sglang.launch_server \\\n    --model-path ./m \\\n    --reasoning-parser=deepseek-r1\n", "deepseek-r1"},
+		{"commands disagree", "sglang.launch_server --reasoning-parser qwen3\nsglang.launch_server --reasoning-parser deepseek-r1\n", ""},
+		{"no parser", "python -m sglang.launch_server --model-path m\n", ""},
+		{"no sglang command", "vllm serve m --reasoning-parser deepseek_r1\n", ""},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			if got := sglangReasoningParser(tt.card); got != tt.want {
+				t.Errorf("sglangReasoningParser() = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}

internal/hf/testdata/README_qwen3_hybrid.md

@@ -0,0 +1,9 @@
+For deployment, you can use `sglang>=0.4.6.post1` or `vllm>=0.8.5` or to create an OpenAI-compatible API endpoint:
+- SGLang:
+    ```shell
+    python -m sglang.launch_server --model-path Qwen/Qwen3-0.6B --reasoning-parser qwen3
+    ```
+- vLLM:
+    ```shell
+    vllm serve Qwen/Qwen3-0.6B --enable-reasoning --reasoning-parser deepseek_r1
+    ```

internal/web/static/models.js

@@ -460,7 +460,10 @@ async function inspect(repo) {
 
     el.flags.hidden = !d.detected_flags;
     if (d.detected_flags) {
-      el.flags.textContent = `Serving flags found on the model card and accepted by your vLLM: ${d.detected_flags}`;
+      el.flags.textContent = [
+        `Serving flags found on the model card and accepted by your vLLM: ${d.detected_flags}`,
+        ...(d.flag_notes ?? []),
+      ].join("\n");
     }
     el.dropped.hidden = !d.dropped_flags?.length;
     if (d.dropped_flags?.length) {