From 383f12bab51bff6796f867f234ddc93e9a75def7 Mon Sep 17 00:00:00 2001
From: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com>
Date: Sat, 19 Sep 2026 15:19:18 -0700
Subject: [PATCH] Allow llama.cpp --reasoning flag in runtime flag allowlist

Motivation:
llama.cpp's llama-server exposes a --reasoning [on|off|auto] flag that
toggles reasoning/thinking mode in the chat template, but Docker Model
Runner rejects it because it is missing from LlamaCppAllowedFlags. The
map already allows the related --reasoning-format and --reasoning-budget
flags. Without --reasoning, the only way to pin a hybrid thinking model
non-thinking is --chat-template-kwargs '{"enable_thinking": false}',
which current llama.cpp builds flag as deprecated in favor of
--reasoning on / --reasoning off.

Approach:
Add "--reasoning" to the LlamaCppAllowedFlags map in
pkg/inference/runtime_flags_allowlist.go, following the same enum-flag
pattern already used for --pooling and --reasoning-format (no argument
value is validated beyond the existing path-safety and flag-injection
checks, which apply uniformly to all allowed flags). Added a
"reasoning" category to the table-driven allowlist category test and a
new scenario case in the runtime flag validation test.

Validation:
- go build ./pkg/inference/...
- go vet ./pkg/inference/...
- gofmt -l on the changed files (no output)
- golangci-lint run ./pkg/inference/... (0 issues)
- go test ./pkg/inference/... and go test -race ./pkg/inference/... (all pass)
- Confirmed the new test is a genuine regression test: with only the
  allowlist map entry reverted, the new "llama.cpp: reasoning flag
  allowed" case fails with `runtime flag "--reasoning" is not allowed
  for backend "llama.cpp"`; restoring the entry makes it pass.

Report: https://github.com/docker/model-runner/issues/1064
Signed-off-by: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com>
Assisted-by: claude-sonnet-5 (via Claude Code)
---
 pkg/inference/runtime_flags_allowlist.go      | 1 +
 pkg/inference/runtime_flags_allowlist_test.go | 3 +++
 pkg/inference/runtime_flags_test.go           | 7 +++++++
 3 files changed, 11 insertions(+)

diff --git a/pkg/inference/runtime_flags_allowlist.go b/pkg/inference/runtime_flags_allowlist.go
index 7b3d7d3a8..e92dc63c5 100644
--- a/pkg/inference/runtime_flags_allowlist.go
+++ b/pkg/inference/runtime_flags_allowlist.go
@@ -171,6 +171,7 @@ var LlamaCppAllowedFlags = map[string]bool{
 	"--chat-template-kwargs": true,
 	"--jinja":                true, "--no-jinja": true,
 	"--pooling":              true,
+	"--reasoning":            true,
 	"--reasoning-format":     true,
 	"--reasoning-budget":     true,
 	"--prefill-assistant":    true,
diff --git a/pkg/inference/runtime_flags_allowlist_test.go b/pkg/inference/runtime_flags_allowlist_test.go
index a73ec3890..3e22f3e92 100644
--- a/pkg/inference/runtime_flags_allowlist_test.go
+++ b/pkg/inference/runtime_flags_allowlist_test.go
@@ -161,6 +161,9 @@ func TestLlamaCppAllowedFlags_Categories(t *testing.T) {
 			"--embeddings", "--embedding", "--reranking", "--rerank",
 			"--metrics", "--no-metrics", "--jinja", "--no-jinja",
 		},
+		"reasoning": {
+			"--reasoning", "--reasoning-format", "--reasoning-budget",
+		},
 		"speculative": {
 			"--spec-draft-n-max", "--draft-n-max", "--draft-max", "--spec-draft-n-min", "--draft-n-min", "--draft-min",
 			"--spec-draft-p-min", "--draft-p-min",
diff --git a/pkg/inference/runtime_flags_test.go b/pkg/inference/runtime_flags_test.go
index 78fb84a39..d0fce6b4a 100644
--- a/pkg/inference/runtime_flags_test.go
+++ b/pkg/inference/runtime_flags_test.go
@@ -91,6 +91,13 @@ func TestValidateRuntimeFlags(t *testing.T) {
 			expectError: false,
 			description: "Sampling flags should be allowed",
 		},
+		{
+			name:        "llama.cpp: reasoning flag allowed",
+			backend:     "llama.cpp",
+			flags:       []string{"--reasoning", "off"},
+			expectError: false,
+			description: "Reasoning flag should be allowed",
+		},
 		{
 			name:    "llama.cpp: real-world flags from issue 515",
 			backend: "llama.cpp",