Studio: an absent speculative mode is auto, not off, for the GPU arbiter

_canonicalize_spec_mode returns None for None, empty and whitespace alike, and
every consumer resolves that to auto (the module docstring says so, and
load_model spells it '_canonicalize_spec_mode(...) or auto'), where a remote or
local sidecar can be discovered and launched with its default GPU offload.

zero_vram_chat_load grouped None with off, so a manual gpu_layers=0 load that
omitted the mode was classified zero-VRAM and skipped acquire_for(CHAT). The
launch-time mask disagreed: it reads the FINISHED argv, sees the drafter that
auto had since added, and keeps the GPUs visible. Same load, no arbiter, real
VRAM, free to land on a resident image or video pipeline and OOM both. Only the
canonical off is speculation-free.

The three existing tests passed the mode positionally as absent while meaning
off, so they encoded the assumption being corrected; they now say off
explicitly. The empty-string case moves with them: it canonicalizes to None like
an omitted mode, so it was never off either.

studio/backend/core/inference/llama_cpp.py

@@ -1800,8 +1800,17 @@ def zero_vram_chat_load(
         return False
     # Any speculative mode may launch a GPU drafter and only the request's own knobs are known here, so treat every selection as GPU-bearing except
     # "off". Canonicalize first: the UI sends the literal "off", which a bare truthiness test read as "speculation requested".
-    spec_mode = _canonicalize_spec_mode(speculative_type)
-    if needs_mmproj or spec_mode not in (None, "off"):
+    # ...and an ABSENT mode is not a quiet "off": _canonicalize_spec_mode returns None
+    # for None, "" and whitespace alike, and every consumer resolves that to "auto"
+    # (`_canonicalize_spec_mode(...) or "auto"` in load_model), where a remote or local
+    # sidecar can be discovered and launched with its default GPU offload. Grouping None
+    # with "off" here let an API or defaulted load skip acquire_for(CHAT) while the
+    # launch-time mask, which reads the FINISHED argv and so sees the drafter that was
+    # added, kept the GPUs visible. The two then disagreed about the same load: no
+    # arbiter, real VRAM, free to land on a resident image or video pipeline and OOM
+    # both. Only the canonical "off" is speculation-free.
+    spec_mode = _canonicalize_spec_mode(speculative_type) or "auto"
+    if needs_mmproj or spec_mode != "off":
         return False
     if LlamaCppBackend._is_vulkan_backend():
         return False

studio/backend/tests/test_gpu_memory_mode.py

@@ -1380,25 +1380,29 @@ def test_zero_vram_chat_load_only_for_a_deliberate_cpu_only_offload(not_vulkan):
     # Manual + gpu_layers=0 is the one shape that launches with the GPUs hidden from the child, so it is the one shape allowed
     # to skip the GPU arbiter. Auto (or any pinned layer count) puts the model on the GPU and must still evict a pipeline.
     zero = llama_cpp_module.zero_vram_chat_load
-    assert zero("manual", 0) is True
-    assert zero("auto", 0) is False
-    assert zero("manual", 1) is False
-    assert zero("manual", -1) is False
+    # Speculation named explicitly off: an ABSENT mode resolves to auto everywhere else
+    # in the module, so it is GPU-bearing and covered by its own test below.
+    assert zero("manual", 0, [], False, "off") is True
+    assert zero("auto", 0, [], False, "off") is False
+    assert zero("manual", 1, [], False, "off") is False
+    assert zero("manual", -1, [], False, "off") is False
 
 
 def test_zero_vram_chat_load_refuses_every_gpu_companion(not_vulkan):
     # The launch-time mask keeps the GPUs visible for a device pin, tensor mode, an mmproj or a drafter, so those loads DO hold
     # VRAM. --mmproj and --model-draft are added by the backend, so their intent arrives as flags.
     zero = llama_cpp_module.zero_vram_chat_load
-    assert zero("manual", 0, ["--device", "CUDA0"]) is False
-    assert zero("manual", 0, ["-dev", "CUDA0"]) is False
-    assert zero("manual", 0, ["--split-mode", "tensor"]) is False
-    assert zero("manual", 0, ["--model-draft", "/tmp/draft.gguf"]) is False
-    assert zero("manual", 0, [], True) is False
+    assert zero("manual", 0, ["--device", "CUDA0"], False, "off") is False
+    assert zero("manual", 0, ["-dev", "CUDA0"], False, "off") is False
+    assert zero("manual", 0, ["--split-mode", "tensor"], False, "off") is False
+    assert zero("manual", 0, ["--model-draft", "/tmp/draft.gguf"], False, "off") is False
+    assert zero("manual", 0, [], True, "off") is False
     assert zero("manual", 0, [], False, "model") is False
     # A CPU-pinned device and a CPU-forced drafter keep it zero-VRAM.
-    assert zero("manual", 0, ["--device", "none"]) is True
-    assert zero("manual", 0, ["--model-draft", "/tmp/d.gguf", "--spec-draft-ngl", "0"]) is True
+    assert zero("manual", 0, ["--device", "none"], False, "off") is True
+    assert zero(
+        "manual", 0, ["--model-draft", "/tmp/d.gguf", "--spec-draft-ngl", "0"], False, "off"
+    ) is True
 
 
 def test_zero_vram_chat_load_exempts_disabled_speculation(not_vulkan):
@@ -1407,7 +1411,9 @@ def test_zero_vram_chat_load_exempts_disabled_speculation(not_vulkan):
     zero = llama_cpp_module.zero_vram_chat_load
     assert zero("manual", 0, [], False, "off") is True
     assert zero("manual", 0, [], False, " OFF ") is True
-    assert zero("manual", 0, [], False, "") is True
+    # Empty is NOT off: it canonicalizes to None like an omitted mode, and every
+    # consumer resolves that to auto, which may launch a drafter on the GPU.
+    assert zero("manual", 0, [], False, "") is False
     assert zero("manual", 0, [], False, "auto") is False
     assert zero("manual", 0, [], False, "mtp") is False
     assert zero("manual", 0, [], False, "default") is False
@@ -1465,3 +1471,20 @@ def test_cmd_companion_ignores_a_projector_pinned_off_the_gpu():
     # llama.cpp assigns rather than accumulates for this one, so the last flag wins.
     cmd = ["llama-server", "--mmproj", "p.gguf", "--no-mmproj-offload", "--mmproj-offload"]
     assert has(cmd, {}) is True
+
+
+def test_zero_vram_chat_load_treats_an_absent_mode_as_auto(not_vulkan):
+    # _canonicalize_spec_mode returns None for None, "" and whitespace alike, and every
+    # consumer resolves that to "auto" (`... or "auto"` in load_model), where a remote or
+    # local sidecar can be discovered and launched with its default GPU offload. Reading
+    # an absent mode as "off" here let an API or defaulted load skip acquire_for(CHAT)
+    # while the launch-time mask, which reads the finished argv and so sees the drafter
+    # that got added, kept the GPUs visible: no arbiter, real VRAM, free to land on a
+    # resident image/video pipeline and OOM both.
+    zero = llama_cpp_module.zero_vram_chat_load
+    assert zero("manual", 0) is False
+    assert zero("manual", 0, [], False, None) is False
+    assert zero("manual", 0, [], False, "") is False
+    assert zero("manual", 0, [], False, "   ") is False
+    # Only the canonical spelling still exempts it.
+    assert zero("manual", 0, [], False, "off") is True