[Code sync] Fix registration of some ops in grok & Fix oss sync scripts (#13990)

Co-authored-by: Stefan He <hebiaobuaa@gmail.com>

python/sglang/srt/environ.py

@@ -166,6 +166,7 @@ class Envs:
     SGLANG_MIN_NEW_TOKEN_RATIO_FACTOR = EnvFloat(0.14)
     SGLANG_NEW_TOKEN_RATIO_DECAY_STEPS = EnvInt(600)
     SGLANG_RETRACT_DECODE_STEPS = EnvInt(20)
+    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION = EnvInt(4096)
 
     # Scheduler: others:
     SGLANG_EMPTY_CACHE_INTERVAL = EnvFloat(-1)  # in seconds. Set if you observe high memory accumulation over a long serving period.

python/sglang/srt/layers/elementwise.py

@@ -1,10 +1,10 @@
-from typing import Tuple
+from typing import Optional, Tuple
 
 import torch
 import triton
 import triton.language as tl
 
-from sglang.srt.utils import is_hip
+from sglang.srt.utils import direct_register_custom_op, is_hip
 
 _is_hip = is_hip()
 
@@ -358,7 +358,11 @@ def experts_combine_kernel(
     tl.store(out_hidden_states + start_index_mlp + offsets, combined_x, mask=mask)
 
 
-def experts_combine_triton(moe_hidden_states, mlp_hidden_states, output_buffer=None):
+def experts_combine_triton(
+    moe_hidden_states: torch.Tensor,
+    mlp_hidden_states: torch.Tensor,
+    output_buffer: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
     assert moe_hidden_states.is_contiguous()
     assert mlp_hidden_states.is_contiguous()
 
@@ -393,9 +397,26 @@ def experts_combine_triton(moe_hidden_states, mlp_hidden_states, output_buffer=N
         hidden_dim,
         **config,
     )
+
     return out_hidden_states
 
 
+def experts_combine_triton_fake(
+    moe_hidden_states: torch.Tensor,
+    mlp_hidden_states: torch.Tensor,
+    output_buffer: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    return torch.empty_like(mlp_hidden_states)
+
+
+direct_register_custom_op(
+    op_name="experts_combine_triton",
+    op_func=experts_combine_triton,
+    mutates_args=[],
+    fake_impl=experts_combine_triton_fake,
+)
+
+
 # gelu on first half of vector
 @triton.jit
 def gelu_and_mul_kernel(

python/sglang/srt/layers/moe/topk.py

@@ -53,6 +53,7 @@ from sglang.srt.utils import (
     is_hip,
     is_npu,
 )
+from sglang.srt.utils.patch_torch import register_fake_if_exists
 
 if TYPE_CHECKING:
     from sglang.srt.layers.quantization import QuantizationConfig
@@ -72,30 +73,12 @@ _is_npu = is_npu()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _is_cuda:
-    from sgl_kernel import kimi_k2_moe_fused_gate, moe_fused_gate
-
-    @torch.library.register_fake("sgl_kernel::kimi_k2_moe_fused_gate")
-    def _kimi_k2_moe_fused_gate(
-        input_tensor,
-        bias,
-        topk,
-        renormalize,
-        routed_scaling_factor,
-        apply_routed_scaling_factor_on_output,
-    ):
-        num_rows = input_tensor.shape[0]
-        topk_weights = input_tensor.new_empty(
-            num_rows,
-            topk,
-            dtype=torch.float32,
-        )
-        topk_ids = input_tensor.new_empty(
-            num_rows,
-            topk,
-            dtype=torch.int32,
-        )
-        return topk_weights, topk_ids
+    from sgl_kernel import moe_fused_gate
 
+    try:
+        from sgl_kernel import kimi_k2_moe_fused_gate
+    except ImportError as e:
+        pass
 
 if _is_cuda or _is_hip:
     from sgl_kernel import topk_softmax
@@ -1044,7 +1027,7 @@ def select_experts(
 if _is_cuda:
 
     @torch.library.register_fake("sgl_kernel::moe_fused_gate")
-    def _(
+    def _moe_fused_gate(
         input_tensor,
         bias,
         num_expert_group,
@@ -1062,3 +1045,25 @@ if _is_cuda:
             (num_rows, topk), dtype=torch.int32, device=input_tensor.device
         )
         return topk_weights, topk_ids
+
+    @register_fake_if_exists("sgl_kernel::kimi_k2_moe_fused_gate")
+    def _kimi_k2_moe_fused_gate(
+        input_tensor,
+        bias,
+        topk,
+        renormalize,
+        routed_scaling_factor,
+        apply_routed_scaling_factor_on_output,
+    ):
+        num_rows = input_tensor.shape[0]
+        topk_weights = input_tensor.new_empty(
+            num_rows,
+            topk,
+            dtype=torch.float32,
+        )
+        topk_ids = input_tensor.new_empty(
+            num_rows,
+            topk,
+            dtype=torch.int32,
+        )
+        return topk_weights, topk_ids

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -49,6 +49,7 @@ from sglang.srt.utils.common import (
     is_sm120_supported,
     next_power_of_2,
 )
+from sglang.srt.utils.patch_torch import register_fake_if_exists
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
@@ -127,7 +128,7 @@ def _sglang_fp4_gemm_fake(
 
 if is_cuda() and (not is_sm120_supported()) and (fp4_quantize is not None):
 
-    @torch.library.register_fake("sgl_kernel::scaled_fp4_quant")
+    @register_fake_if_exists("sgl_kernel::scaled_fp4_quant")
     def _sgl_kernel_scaled_fp4_quant_fake(
         output, input, output_scale, input_global_scale
     ):

python/sglang/srt/utils/common.py

@@ -2753,6 +2753,21 @@ def load_json_config(data: str):
 
 
 def dispose_tensor(x: torch.Tensor):
+    """
+    Dispose a tensor by freeing its memory.
+    During piecewise CUDA graph capture/replay, we skip disposal to avoid
+    interfering with torch.compile's memory tracking and graph recording.
+    """
+
+    # Skip disposal during piecewise CUDA graph to avoid torch.compile issues
+    # we do local import to avoid circular import
+    from sglang.srt.compilation.piecewise_context_manager import (
+        is_in_piecewise_cuda_graph,
+    )
+
+    if is_in_piecewise_cuda_graph():
+        return
+
     x.set_(torch.empty((0,), device=x.device, dtype=x.dtype))