diff --git a/python/sglang/multimodal_gen/configs/sample/sensenova_u1.py b/python/sglang/multimodal_gen/configs/sample/sensenova_u1.py index 9add7ace23..8afe673d86 100644 --- a/python/sglang/multimodal_gen/configs/sample/sensenova_u1.py +++ b/python/sglang/multimodal_gen/configs/sample/sensenova_u1.py @@ -33,6 +33,10 @@ _PUBLIC_OVERRIDE_FIELDS = { "output_quality", "output_compression", "quality", + "profile", + "profile_all_stages", + "num_profiled_timesteps", + "perf_dump_path", } diff --git a/python/sglang/multimodal_gen/runtime/models/sensenova_u1/neo_unify/modeling_qwen3.py b/python/sglang/multimodal_gen/runtime/models/sensenova_u1/neo_unify/modeling_qwen3.py index bb5c79a826..289fd181c1 100644 --- a/python/sglang/multimodal_gen/runtime/models/sensenova_u1/neo_unify/modeling_qwen3.py +++ b/python/sglang/multimodal_gen/runtime/models/sensenova_u1/neo_unify/modeling_qwen3.py @@ -429,6 +429,42 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. """ + if ( + q.is_cuda + and q.dtype is torch.bfloat16 + and k.dtype is torch.bfloat16 + and cos.dtype is torch.bfloat16 + and sin.dtype is torch.bfloat16 + and q.device == k.device == cos.device == sin.device + and not torch.is_grad_enabled() + and not torch.compiler.is_compiling() + and unsqueeze_dim == 1 + and q.ndim == k.ndim == 4 + and q.numel() > 0 + and k.numel() > 0 + and q.shape[2] >= 128 + and q.shape[0] == k.shape[0] + and q.shape[2:] == k.shape[2:] + and q.shape[-1] in (32, 64) + and cos.shape == sin.shape == (q.shape[0], q.shape[2], q.shape[3]) + ): + from sglang.kernels.ops.diffusion.rope.rope_rotate_half_bitexact import ( + fused_rope_rotate_half_bitexact, + ) + + # The kernel rounds both products to bf16 before adding, matching + # the eager chain. Spatial axes are slices of the normalized HW + # heads; materialize those slices in the kernel's BSHD layout. + cos_rows = cos.reshape(-1, cos.shape[-1]).contiguous() + sin_rows = sin.reshape(-1, sin.shape[-1]).contiguous() + q_embed = fused_rope_rotate_half_bitexact( + q.transpose(1, 2).contiguous(), cos_rows, sin_rows + ) + k_embed = fused_rope_rotate_half_bitexact( + k.transpose(1, 2).contiguous(), cos_rows, sin_rows + ) + return q_embed.transpose(1, 2), k_embed.transpose(1, 2) + cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) q_embed = (q * cos) + (rotate_half(q) * sin) diff --git a/python/sglang/multimodal_gen/test/unit/test_sensenova_u1.py b/python/sglang/multimodal_gen/test/unit/test_sensenova_u1.py index f96b9f0a44..917da42be8 100644 --- a/python/sglang/multimodal_gen/test/unit/test_sensenova_u1.py +++ b/python/sglang/multimodal_gen/test/unit/test_sensenova_u1.py @@ -1042,6 +1042,10 @@ def test_sensenova_u1_cli_args_expose_only_sglang_compatible_fields(): guidance_scale=4.5, num_inference_steps=30, num_outputs_per_prompt=2, + profile=True, + profile_all_stages=True, + num_profiled_timesteps=3, + perf_dump_path="/tmp/sensenova-perf.json", cfg_norm="global", timestep_shift=9.0, think_mode=True, @@ -1058,6 +1062,11 @@ def test_sensenova_u1_cli_args_expose_only_sglang_compatible_fields(): assert "cfg_norm" not in cli_args assert "timestep_shift" not in cli_args assert "think_mode" not in cli_args + request = Req(sampling_params=SenseNovaU1SamplingParams(**cli_args)) + assert request.profile + assert request.profile_all_stages + assert request.num_profiled_timesteps == 3 + assert request.perf_dump_path == "/tmp/sensenova-perf.json" def test_sensenova_u1_generation_stage_uses_sglang_params_and_single_model_batch():