diffusion: Fix silent override of explicit guidance_scale=0 (#4999)

Signed-off-by: Harenome Ranaivoarivony-Razanajato <harenome.razanajato1@huawei.com>

tests/diffusion/test_diffusion_request.py

@@ -72,3 +72,40 @@ def test_tp_seed_same_across_ranks_and_varies_across_requests():
 
     # Seeds must vary across requests (non-determinism preserved).
     assert len(set(seeds)) == n_requests, f"Expected {n_requests} unique seeds but got {len(set(seeds))}: {seeds}"
+
+
+def _request_with_guidance(**sampling_kwargs) -> OmniDiffusionRequest:
+    return OmniDiffusionRequest(
+        prompt={"prompt": "a cup of coffee on a table"},
+        sampling_params=OmniDiffusionSamplingParams(num_inference_steps=1, **sampling_kwargs),
+        request_id="request-test",
+    )
+
+
+def test_omitted_guidance_scale_is_unset_and_resolved_to_one():
+    # No guidance_scale passed: the caller did not set it, so it must NOT count as
+    # "provided" (each pipeline applies its own default), and __post_init__ resolves the
+    # None "unset" sentinel to a concrete float so no consumer ever observes None.
+    req = _request_with_guidance()
+
+    assert req.sampling_params.guidance_scale_provided is False
+    assert req.sampling_params.guidance_scale == 1.0
+
+
+def test_explicit_zero_guidance_scale_is_honored():
+    # Regression (#4998): an explicit guidance_scale=0 (disable CFG) must be honored,
+    # not mistaken for "unset" and overridden by the pipeline's model default. This is
+    # the sentinel collision that the None default + ``is not None`` check fixes.
+    req = _request_with_guidance(guidance_scale=0.0)
+
+    assert req.sampling_params.guidance_scale_provided is True
+    assert req.sampling_params.guidance_scale == 0.0
+
+
+@pytest.mark.parametrize("value", [0.999, 1.0, 5.0])
+def test_explicit_nonzero_guidance_scale_is_honored(value):
+    # Any explicitly-set value (truthy or not) is honored as provided and preserved.
+    req = _request_with_guidance(guidance_scale=value)
+
+    assert req.sampling_params.guidance_scale_provided is True
+    assert req.sampling_params.guidance_scale == value

vllm_omni/diffusion/models/flux/pipeline_flux_kontext.py

@@ -640,7 +640,9 @@ class FluxKontextPipeline(
         width = req.sampling_params.width or width
         num_inference_steps = req.sampling_params.num_inference_steps or num_inference_steps
         sigmas = req.sampling_params.sigmas or sigmas
-        guidance_scale = req.sampling_params.guidance_scale or guidance_scale
+        guidance_scale = (
+            req.sampling_params.guidance_scale if req.sampling_params.guidance_scale_provided else guidance_scale
+        )
         num_images_per_prompt = (
             req.sampling_params.num_outputs_per_prompt
             if req.sampling_params.num_outputs_per_prompt > 0

vllm_omni/diffusion/models/glm_image/pipeline_glm_image.py

@@ -733,7 +733,7 @@ class GlmImagePipeline(nn.Module, DiffusionPipelineProfilerMixin, SupportsCompon
         )
         width = ar_width or req.sampling_params.width or img_width or self.default_sample_size * self.vae_scale_factor
         num_inference_steps = req.sampling_params.num_inference_steps or 50
-        guidance_scale = req.sampling_params.guidance_scale or 1.5
+        guidance_scale = req.sampling_params.guidance_scale if req.sampling_params.guidance_scale_provided else 1.5
 
         # Ensure dimensions are multiples of vae_scale_factor * patch_size
         multiple_of = self.vae_scale_factor * self._patch_size

vllm_omni/diffusion/models/soulx_singer/pipeline_soulx_singer_base.py

@@ -353,7 +353,7 @@ class FlowMatchingAudioPipeline(
             sampling_params.extra_args = extra_args
 
         num_inference_steps = sampling_params.num_inference_steps or 32
-        guidance_scale = sampling_params.guidance_scale or 3.0
+        guidance_scale = sampling_params.guidance_scale if sampling_params.guidance_scale_provided else 3.0
         generator = self._resolve_diffusion_generator(sampling_params)
 
         with self._stage_timer("consume_payload"):

vllm_omni/diffusion/request.py

@@ -42,11 +42,11 @@ class OmniDiffusionRequest:
             self.sampling_params.seed = random.randint(0, 2**31 - 1)
 
         # Detect whether user explicitly provided guidance_scale.
-        # The sentinel default is 0.0 (false-like); any truthy value means
-        # the caller set it intentionally.  We must resolve this BEFORE
-        # auto-filling guidance_scale_2, otherwise the sentinel leaks into
-        # guidance_scale_2.
-        if self.sampling_params.guidance_scale:
+        # The unset sentinel is None (see OmniDiffusionSamplingParams); any value
+        # the caller set -- including an explicit 0.0 (disable CFG) -- must be
+        # honored. Resolve before auto-filling guidance_scale_2, else the sentinel
+        # leaks into guidance_scale_2.
+        if self.sampling_params.guidance_scale is not None:
             self.sampling_params.guidance_scale_provided = True
         else:
             self.sampling_params.guidance_scale = 1.0

vllm_omni/inputs/data.py

@@ -263,7 +263,7 @@ class OmniDiffusionSamplingParams:
     # Scheduler parameters – ``None`` means "not explicitly set by the caller";
     # each pipeline's ``forward()`` decides its own model-specific default.
     num_inference_steps: int | None = None
-    guidance_scale: float = 0.0
+    guidance_scale: float | None = None
     guidance_scale_provided: bool = False
     guidance_scale_2: float | None = None
     guidance_scale_2_provided: bool = False