[Diffusion]Skipping tensor copying for non-BCG GLM-Image workflows (#33688)

Co-authored-by: Elizaveta Martirosian <elizaveta.martirosian@gmail.com>
Co-authored-by: ronnie_zheng <zl19940307@163.com>

python/sglang/multimodal_gen/runtime/models/dits/glm_image.py

@@ -56,6 +56,9 @@ from sglang.multimodal_gen.runtime.platforms import (
     current_platform,
 )
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+from sglang.srt.model_executor.runner_backend_utils.breakable_cuda_graph import (
+    is_in_breakable_cuda_graph,
+)
 
 logger = init_logger(__name__)
 
@@ -977,8 +980,15 @@ class GlmImageTransformer2DModel(CachableDiT, LayerwiseOffloadableModuleMixin):
 
         hidden_states = self.image_projector(hidden_states)
         encoder_hidden_states = self.glyph_projector(encoder_hidden_states)
+
         prior_embedding = self.prior_token_embedding(prior_token_id)
-        prior_embedding = prior_embedding.masked_fill(prior_token_drop.unsqueeze(-1), 0)
+        if is_in_breakable_cuda_graph():
+            prior_embedding = prior_embedding.masked_fill(
+                prior_token_drop.unsqueeze(-1), 0
+            )
+        else:
+            prior_embedding[prior_token_drop] *= 0.0
+
         prior_hidden_states = self.prior_projector(prior_embedding)
         # SP: when latents are H-sharded, hidden_states has fewer patches than prior_hidden_states.
         # Shard prior_hidden_states along seq dim to match (prior is row-major, same as latent patches).

python/sglang/multimodal_gen/test/scripts/gen_perf_baselines.py

@@ -8,6 +8,7 @@ from pathlib import Path
 
 from openai import OpenAI
 
+from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.test.server.test_server_utils import (
     ServerManager,
     get_generate_fn,
@@ -23,6 +24,9 @@ from sglang.multimodal_gen.test.test_utils import (
 
 
 def _all_cases() -> list[DiffusionTestCase]:
+    if current_platform.is_npu():
+        import sglang.multimodal_gen.test.server.ascend.testcase_configs_npu as cfg
+    else:
         import sglang.multimodal_gen.test.server.testcase_configs as cfg
 
     cases: list[DiffusionTestCase] = []

python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json

@@ -395,7 +395,7 @@
                 "GlmImageAR": 69249.94,
                 "GlmImageBeforeDenoisingStage": 61.78,
                 "DenoisingStage": 17750.92,
-                "DecodingStage": 795.68
+                "DecodingStage": 277.36
             },
             "denoise_step_ms": {
                 "0": 272.46,