Fix Inkling kernel imports after migration (#32076)

python/sglang/srt/models/inkling_common/attn.py

@@ -364,7 +364,8 @@ class InklingAttention(nn.Module):
 
     def _fused_attn_prologue_verify(self, q, k, v, forward_batch, log_scaling_tau=None):
         """Fused target-verify {k/v sconv + save_windows + qk-norm (+ KV store)}
-        (jit_kernel/inkling_attn_prologue.py); returns ``(q, k, v, did_store)``.
+        (kernels/ops/model/inkling/inkling_attn_prologue.py); returns
+        ``(q, k, v, did_store)``.
 
         The fused kernel writes raw bf16 KV, so it only does the store when the
         KV pool is bf16: full layers at ``out_cache_loc`` in the full pool,

python/sglang/srt/models/inkling_common/kernels/comm.py

@@ -94,7 +94,7 @@ def _ar_jit():
     lazy so importing comm.py doesn't pull in the JIT machinery)."""
     if not is_cuda():
         return None
-    from sglang.jit_kernel import inkling_all_reduce
+    from sglang.kernels.ops.model.inkling import inkling_all_reduce
 
     return inkling_all_reduce
 
@@ -103,7 +103,7 @@ def _ar_jit():
 def _ar_fused_jit():
     if not is_cuda():
         return None
-    from sglang.jit_kernel import inkling_ar_fused
+    from sglang.kernels.ops.model.inkling import inkling_ar_fused
 
     return inkling_ar_fused
 
@@ -240,7 +240,8 @@ def ar_sconv_norm_fusable(
     """True when a decode {all-reduce -> sconv -> add+RMSNorm} chain
     (attn-side: wo_ud AR -> attn_sconv -> mlp_norm; MoE-side: MoE AR ->
     mlp_sconv -> next attn_norm)
-    can run as the single fused kernel (jit_kernel/inkling_ar_fused.py). Must be
+    can run as the single fused kernel
+    (kernels/ops/model/inkling/inkling_ar_fused.py). Must be
     evaluated identically by the producing layer (MoE ``reduce=False``) and the
     consuming layer/tail -- it is a pure function of per-forward state."""
     if not is_cuda():
@@ -674,7 +675,7 @@ def all_gather_hidden(input: torch.Tensor, group: GroupCoordinator) -> torch.Ten
 def _ar_ssconv_jit():
     if not is_cuda():
         return None
-    from sglang.jit_kernel import inkling_ar_scattered_sconv
+    from sglang.kernels.ops.model.inkling import inkling_ar_scattered_sconv
 
     return inkling_ar_scattered_sconv
 
@@ -688,7 +689,8 @@ def scattered_ar_sconv_fusable(
 ) -> bool:
     """True when an extend {reduce_scatter_hidden -> sconv(shard) ->
     all_gather_hidden} chain can run as the single fused v3/v3b-style kernel
-    (jit_kernel/inkling_ar_scattered_sconv.py). Pure function of per-forward
+    (kernels/ops/model/inkling/inkling_ar_scattered_sconv.py). Pure function of
+    per-forward
     state -- the producing layer (reduce=False) and the consuming site must
     evaluate it identically."""
     if not is_cuda():