[diffusion] fix: fix Flux2-Klein prompt tokenization length to 512 and add regression coverage (#21407)

python/sglang/multimodal_gen/configs/pipeline_configs/flux.py

@@ -435,6 +435,17 @@ class Flux2PipelineConfig(FluxPipelineConfig):
         default_factory=lambda: (flux2_postprocess_text,)
     )
     vae_config: VAEConfig = field(default_factory=Flux2VAEConfig)
+    text_encoder_extra_args: list[dict] = field(
+        default_factory=lambda: [
+            dict(
+                max_length=512,
+                padding="max_length",
+                truncation=True,
+                return_overflowing_tokens=False,
+                return_length=False,
+            )
+        ]
+    )
 
     def tokenize_prompt(self, prompts: list[str], tokenizer, tok_kwargs) -> dict:
         # flatten to 1-d list
@@ -679,7 +690,9 @@ class Flux2KleinPipelineConfig(Flux2PipelineConfig):
         texts = [_apply_chat_template(prompt) for prompt in prompts]
 
         tok_kwargs = dict(tok_kwargs or {})
-        max_length = tok_kwargs.pop("max_length", 512)
+        tok_kwargs.pop("max_length", None)
+        # Flux2 Klein uses max_length 512.
+        max_length = 512
         padding = tok_kwargs.pop("padding", "max_length")
         truncation = tok_kwargs.pop("truncation", True)
         return_tensors = tok_kwargs.pop("return_tensors", "pt")