sglang 0.4.6.post5__py3-none-any.whl → 0.4.7__py3-none-any.whl

sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -28,8 +28,9 @@ else:
 import logging
 
 _is_hip = is_hip()
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
-if _is_hip:
+if _use_aiter:
     from aiter import ActivationType
     from aiter.fused_moe_bf16_asm import ck_moe_2stages
     from aiter.ops.shuffle import shuffle_weight
@@ -104,7 +105,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         set_weight_attrs(w2_weight, extra_weight_attrs)
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
-        if _is_hip and get_bool_env_var("SGLANG_AITER_MOE"):
+        if _use_aiter:
             layer.w13_weight = torch.nn.Parameter(
                 shuffle_weight(layer.w13_weight.data, (16, 16)),
                 requires_grad=False,
@@ -127,6 +128,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         use_grouped_topk: bool,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
@@ -144,6 +146,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             use_grouped_topk=use_grouped_topk,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
+            num_fused_shared_experts=num_fused_shared_experts,
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
             activation=activation,
@@ -163,6 +166,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         renormalize: bool,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
@@ -179,12 +183,13 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             renormalize=renormalize,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
+            num_fused_shared_experts=num_fused_shared_experts,
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
             routed_scaling_factor=routed_scaling_factor,
         )
 
-        if _is_hip and get_bool_env_var("SGLANG_AITER_MOE"):
+        if _use_aiter:
             assert not no_combine, "unsupported"
             if apply_router_weight_on_input:
                 assert (
@@ -220,6 +225,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 activation=activation,
                 apply_router_weight_on_input=apply_router_weight_on_input,
                 no_combine=no_combine,
+                routed_scaling_factor=routed_scaling_factor,
             )
 
     def forward_cpu(
@@ -232,6 +238,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         renormalize: bool,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         inplace: bool = True,
@@ -245,6 +252,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             renormalize,
             topk_group,
             num_expert_group,
+            num_fused_shared_experts,
             custom_routing_function,
             correction_bias,
         )
@@ -289,6 +297,7 @@ class FusedMoE(torch.nn.Module):
         renormalize: bool = True,
         use_grouped_topk: bool = False,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         topk_group: Optional[int] = None,
         quant_config: Optional[QuantizationConfig] = None,
         tp_size: Optional[int] = None,
@@ -307,6 +316,7 @@ class FusedMoE(torch.nn.Module):
         if params_dtype is None:
             params_dtype = torch.get_default_dtype()
 
+        self.hidden_size = hidden_size
         self.tp_size = (
             tp_size if tp_size is not None else get_tensor_model_parallel_world_size()
         )
@@ -321,6 +331,7 @@ class FusedMoE(torch.nn.Module):
         if self.use_grouped_topk:
             assert num_expert_group is not None and topk_group is not None
         self.num_expert_group = num_expert_group
+        self.num_fused_shared_experts = num_fused_shared_experts
         self.topk_group = topk_group
         self.custom_routing_function = custom_routing_function
         self.correction_bias = correction_bias
@@ -546,7 +557,8 @@ class FusedMoE(torch.nn.Module):
             loaded_weight = loaded_weight.to(param.data.device)
 
             if (
-                param.data[expert_id] != 1
+                "compressed" in self.quant_method.__class__.__name__.lower()
+                and param.data[expert_id] != 1
                 and (param.data[expert_id] - loaded_weight).abs() > 1e-5
             ):
                 raise ValueError(
@@ -570,6 +582,23 @@ class FusedMoE(torch.nn.Module):
                 tp_rank=tp_rank,
             )
             return
+        if "ModelOpt" in self.quant_method.__class__.__name__:
+            if "weight_scale_2" in weight_name or "input_scale" in weight_name:
+                self._load_per_tensor_weight_scale(
+                    shard_id=shard_id,
+                    param=param,
+                    loaded_weight=loaded_weight,
+                    expert_id=expert_id,
+                )
+            elif "weight" in weight_name:
+                self._load_model_weight_or_group_weight_scale(
+                    shard_id=shard_id,
+                    shard_dim=shard_dim,
+                    loaded_weight=loaded_weight,
+                    expert_data=expert_data,
+                    tp_rank=tp_rank,
+                )
+            return
 
         # Case weight scales and zero_points
         if "scale" in weight_name or "zero" in weight_name:
@@ -651,6 +680,7 @@ class FusedMoE(torch.nn.Module):
             use_grouped_topk=self.use_grouped_topk,
             topk_group=self.topk_group,
             num_expert_group=self.num_expert_group,
+            num_fused_shared_experts=self.num_fused_shared_experts,
             custom_routing_function=self.custom_routing_function,
             correction_bias=self.correction_bias,
             activation=self.activation,

sglang/srt/layers/moe/topk.py

@@ -18,6 +18,7 @@ from typing import Callable, Optional
 import torch
 import torch.nn.functional as F
 
+from sglang.srt.managers import expert_location_dispatch
 from sglang.srt.managers.expert_distribution import (
     ExpertDistributionRecorder,
     get_global_expert_distribution_recorder,
@@ -65,6 +66,7 @@ def fused_topk(
     gating_output: torch.Tensor,
     topk: int,
     renormalize: bool,
+    num_token_non_padded: Optional[torch.Tensor] = None,
     expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
 ):
     assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
@@ -87,9 +89,27 @@ def fused_topk(
     )
     del token_expert_indicies
 
+    return _fused_topk_postprocess(
+        topk_weights=topk_weights,
+        topk_ids=topk_ids,
+        renormalize=renormalize,
+        expert_location_dispatch_info=expert_location_dispatch_info,
+        num_token_non_padded=num_token_non_padded,
+    )
+
+
+@torch.compile(dynamic=True, backend=get_compiler_backend())
+def _fused_topk_postprocess(
+    topk_weights,
+    topk_ids,
+    renormalize,
+    expert_location_dispatch_info,
+    num_token_non_padded,
+):
     if renormalize:
         topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
     topk_ids = topk_ids_logical_to_physical(topk_ids, expert_location_dispatch_info)
+    _mask_topk_ids_padded_region(topk_ids, num_token_non_padded)
     return topk_weights, topk_ids
 
 
@@ -102,7 +122,7 @@ def grouped_topk(
     renormalize: bool,
     num_expert_group: int = 0,
     topk_group: int = 0,
-    n_share_experts_fusion: int = 0,
+    num_fused_shared_experts: int = 0,
     routed_scaling_factor: Optional[float] = None,
     num_token_non_padded: Optional[torch.Tensor] = None,
     expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
@@ -127,10 +147,10 @@ def grouped_topk(
     )  # [n, e]
     tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0)  # [n, e]
     topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
-    if n_share_experts_fusion:
+    if num_fused_shared_experts:
         topk_ids[:, -1] = torch.randint(
             low=num_experts,
-            high=num_experts + n_share_experts_fusion,
+            high=num_experts + num_fused_shared_experts,
             size=(topk_ids.size(0),),
             dtype=topk_ids.dtype,
             device=topk_ids.device,
@@ -140,7 +160,7 @@ def grouped_topk(
     if renormalize:
         topk_weights_sum = (
             topk_weights.sum(dim=-1, keepdim=True)
-            if n_share_experts_fusion == 0
+            if num_fused_shared_experts == 0
             else topk_weights[:, :-1].sum(dim=-1, keepdim=True)
         )
         topk_weights = topk_weights / topk_weights_sum
@@ -159,7 +179,7 @@ def biased_grouped_topk_impl(
     renormalize: bool,
     num_expert_group: int = 0,
     topk_group: int = 0,
-    n_share_experts_fusion: int = 0,
+    num_fused_shared_experts: int = 0,
     routed_scaling_factor: Optional[float] = None,
     num_token_non_padded: Optional[torch.Tensor] = None,
     expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
@@ -191,10 +211,10 @@ def biased_grouped_topk_impl(
     _, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)
     topk_weights = scores.gather(1, topk_ids)
 
-    if n_share_experts_fusion:
+    if num_fused_shared_experts:
         topk_ids[:, -1] = torch.randint(
             low=num_experts,
-            high=num_experts + n_share_experts_fusion,
+            high=num_experts + num_fused_shared_experts,
             size=(topk_ids.size(0),),
             dtype=topk_ids.dtype,
             device=topk_ids.device,
@@ -204,7 +224,7 @@ def biased_grouped_topk_impl(
     if renormalize:
         topk_weights_sum = (
             topk_weights.sum(dim=-1, keepdim=True)
-            if n_share_experts_fusion == 0
+            if num_fused_shared_experts == 0
             else topk_weights[:, :-1].sum(dim=-1, keepdim=True)
         )
         topk_weights = topk_weights / topk_weights_sum
@@ -238,7 +258,7 @@ def biased_grouped_topk(
     num_expert_group: int = 0,
     topk_group: int = 0,
     compiled: bool = True,
-    n_share_experts_fusion: int = 0,
+    num_fused_shared_experts: int = 0,
     routed_scaling_factor: Optional[float] = None,
     num_token_non_padded: Optional[torch.Tensor] = None,
     expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
@@ -246,7 +266,7 @@ def biased_grouped_topk(
     assert (
         routed_scaling_factor is not None
     ), "routed_scaling_factor is required for biased_grouped_topk"
-    # TODO: moe_fused_gate kernel is not supported for n_share_experts_fusion > 0 now.
+    # TODO: moe_fused_gate kernel is not supported for num_fused_shared_experts > 0 now.
     if (
         _is_cuda
         and gating_output.shape[1] // num_expert_group
@@ -259,7 +279,7 @@ def biased_grouped_topk(
             num_expert_group,
             topk_group,
             topk,
-            n_share_experts_fusion,
+            num_fused_shared_experts,
             routed_scaling_factor,
         )
         # TODO merge into kernel for this branch
@@ -287,7 +307,7 @@ def biased_grouped_topk(
             renormalize,
             num_expert_group,
             topk_group,
-            n_share_experts_fusion=n_share_experts_fusion,
+            num_fused_shared_experts=num_fused_shared_experts,
             routed_scaling_factor=routed_scaling_factor,
             num_token_non_padded=num_token_non_padded,
             expert_location_dispatch_info=expert_location_dispatch_info,
@@ -302,6 +322,7 @@ def select_experts(
     renormalize: bool,
     topk_group: Optional[int] = None,
     num_expert_group: Optional[int] = None,
+    num_fused_shared_experts: int = 0,
     custom_routing_function: Optional[Callable] = None,
     correction_bias: Optional[torch.Tensor] = None,
     torch_native: bool = False,
@@ -309,7 +330,14 @@ def select_experts(
     num_token_non_padded: Optional[torch.Tensor] = None,
     expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
 ):
-    n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
+    router_logits, correction_bias = (
+        expert_location_dispatch.transform_select_experts_inputs(
+            router_logits=router_logits,
+            correction_bias=correction_bias,
+            info=expert_location_dispatch_info,
+        )
+    )
+
     # DeepSeek V2/V3/R1 series models use grouped_top_k
     if use_grouped_topk:
         assert topk_group is not None
@@ -322,7 +350,7 @@ def select_experts(
                 renormalize=renormalize,
                 num_expert_group=num_expert_group,
                 topk_group=topk_group,
-                n_share_experts_fusion=n_share_experts_fusion,
+                num_fused_shared_experts=num_fused_shared_experts,
                 routed_scaling_factor=routed_scaling_factor,
                 num_token_non_padded=num_token_non_padded,
                 expert_location_dispatch_info=expert_location_dispatch_info,
@@ -336,7 +364,7 @@ def select_experts(
                 renormalize=renormalize,
                 num_expert_group=num_expert_group,
                 topk_group=topk_group,
-                n_share_experts_fusion=n_share_experts_fusion,
+                num_fused_shared_experts=num_fused_shared_experts,
                 routed_scaling_factor=routed_scaling_factor,
                 num_token_non_padded=num_token_non_padded,
                 expert_location_dispatch_info=expert_location_dispatch_info,
@@ -353,15 +381,13 @@ def select_experts(
             renormalize=renormalize,
         )
     elif custom_routing_function is None:
-        assert (
-            num_token_non_padded is None
-        ), "num_token_non_padded is not yet supported in fused_topk"
         # Qwen3MOE uses fused_topk
         topk_weights, topk_ids = fused_topk(
             hidden_states=hidden_states,
             gating_output=router_logits,
             topk=top_k,
             renormalize=renormalize,
+            num_token_non_padded=num_token_non_padded,
             expert_location_dispatch_info=expert_location_dispatch_info,
         )
     else:

sglang/srt/layers/multimodal.py

@@ -32,8 +32,8 @@ def hash_kernel(
     offsets = block_start + tl.arange(0, BLOCK_SIZE)
     mask = offsets < n_elements
 
-    data = tl.load(input_ptr + offsets, mask=mask, other=0)
-    mixed = data ^ (offsets + XCONST)
+    data = tl.load(input_ptr + offsets, mask=mask, other=0).to(tl.int64)
+    mixed = data ^ (offsets.to(tl.int64) + XCONST)
     hash_val = mixed * PRIME
     hash_val = hash_val ^ (hash_val >> 16)
     hash_val = hash_val * (PRIME ^ XCONST)
@@ -53,7 +53,7 @@ def gpu_tensor_hash(tensor: torch.Tensor) -> int:
     BLOCK_SIZE = 1024
     grid = (triton.cdiv(n, BLOCK_SIZE),)
 
-    intermediate_hashes = torch.empty(n, dtype=torch.int32, device=tensor.device)
+    intermediate_hashes = torch.empty(n, dtype=torch.int64, device=tensor.device)
 
     hash_kernel[grid](
         tensor,

sglang/srt/layers/quantization/__init__.py

@@ -114,7 +114,7 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
     if quantization in VLLM_QUANTIZATION_METHODS and not VLLM_AVAILABLE:
         raise ValueError(
             f"{quantization} quantization requires some operators from vllm. "
-            "Please install vllm by `pip install vllm==0.8.4`"
+            "Please install vllm by `pip install vllm==0.9.0.1`"
         )
 
     return QUANTIZATION_METHODS[quantization]
@@ -289,6 +289,7 @@ def monkey_patch_moe_apply(class_obj: "FusedMoEMethodBase"):
         use_grouped_topk: bool,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
@@ -315,7 +316,7 @@ def monkey_patch_moe_apply(class_obj: "FusedMoEMethodBase"):
         if correction_bias is not None:
             if not has_correction_bias:
                 raise ValueError(
-                    "Please increase the version of your vllm. Try `pip install vllm==0.8.4`"
+                    "Please increase the version of your vllm. Try `pip install vllm==0.9.0.1`"
                 )
             kwargs["e_score_correction_bias"] = correction_bias
         return original_apply(**kwargs)

sglang/srt/layers/quantization/blockwise_int8.py

@@ -367,6 +367,7 @@ class BlockInt8MoEMethod:
         use_grouped_topk: bool,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
@@ -387,6 +388,7 @@ class BlockInt8MoEMethod:
             renormalize=renormalize,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
+            num_fused_shared_experts=num_fused_shared_experts,
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
             routed_scaling_factor=routed_scaling_factor,
@@ -409,4 +411,5 @@ class BlockInt8MoEMethod:
             a2_scale=layer.w2_input_scale,
             block_shape=self.quant_config.weight_block_size,
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )

sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -272,6 +272,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         use_grouped_topk: bool = False,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         global_num_experts: int = -1,
         expert_map: Optional[torch.Tensor] = None,
         custom_routing_function: Optional[Callable] = None,
@@ -294,6 +295,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             renormalize=renormalize,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
+            num_fused_shared_experts=num_fused_shared_experts,
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
             routed_scaling_factor=routed_scaling_factor,
@@ -315,6 +317,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
             apply_router_weight_on_input=apply_router_weight_on_input,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
 
@@ -627,6 +630,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
         use_grouped_topk: bool = False,
         topk_group: Optional[int] = None,
         num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: int = 0,
         global_num_experts: int = -1,
         expert_map: Optional[torch.Tensor] = None,
         custom_routing_function: Optional[Callable] = None,
@@ -651,6 +655,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
             renormalize=renormalize,
             topk_group=topk_group,
             num_expert_group=num_expert_group,
+            num_fused_shared_experts=num_fused_shared_experts,
             custom_routing_function=custom_routing_function,
             scoring_func=scoring_func,
             correction_bias=correction_bias,

sglang/srt/layers/quantization/deep_gemm.py

@@ -17,10 +17,10 @@ _ENABLE_JIT_DEEPGEMM = False
 try:
     import deep_gemm
     from deep_gemm import get_num_sms
+    from deep_gemm.jit import build
     from deep_gemm.jit.compiler import get_nvcc_compiler
     from deep_gemm.jit_kernels.gemm import get_best_configs
     from deep_gemm.jit_kernels.runtime import FP8GemmRuntime, GemmType
-    from deep_gemm.jit_kernels.tuner import jit_tuner
 
     sm_version = get_device_sm()
     if sm_version == 90:
@@ -148,32 +148,28 @@ def _compile_grouped_gemm_nt_f8f8bf16_masked_one(
     block_k = 128
     num_tma_threads = 128
     num_math_threads_per_group = 128
+
     kwargs = {
+        "GEMM_TYPE": GemmType.GroupedMasked,
         "NUM_TMA_THREADS": num_tma_threads,
         "NUM_MATH_THREADS_PER_GROUP": num_math_threads_per_group,
+        "N": n,
+        "K": k,
+        "NUM_GROUPS": 1,
+        "BLOCK_M": block_m,
+        "BLOCK_N": block_n,
         "BLOCK_K": block_k,
+        "SWIZZLE_D_MODE": smem_config[1],
+        "BLOCK_N_PADDING": smem_config[2],
+        "NUM_STAGES": num_stages,
+        "NUM_TMA_MULTICAST": tma_multicast_config[0],
+        "IS_TMA_MULTICAST_ON_A": tma_multicast_config[1],
         "NUM_SMS": num_sms,
         "SMEM_SIZE": smem_config[0],
     }
-    _, _ = jit_tuner.compile_and_tune(
-        name="m_grouped_gemm_fp8_fp8_bf16_nt",
-        keys={
-            "N": n,
-            "K": k,
-            "BLOCK_M": block_m,
-            "BLOCK_N": block_n,
-            "SWIZZLE_D_MODE": smem_config[1],
-            "BLOCK_N_PADDING": smem_config[2],
-            "NUM_GROUPS": num_groups,
-            "NUM_STAGES": num_stages,
-            "NUM_TMA_MULTICAST": tma_multicast_config[0],
-            "IS_TMA_MULTICAST_ON_A": tma_multicast_config[1],
-            "GEMM_TYPE": GemmType.GroupedMasked,
-        },
-        space=(),
-        kwargs=kwargs,
-        runtime_cls=FP8GemmRuntime,
-    )
+
+    code = FP8GemmRuntime.generate(kwargs)
+    _ = build("m_grouped_gemm_fp8_fp8_bf16_nt", code, FP8GemmRuntime, kwargs)
 
 
 def _compile_grouped_gemm_nt_f8f8bf16_contig_one(
@@ -187,31 +183,26 @@ def _compile_grouped_gemm_nt_f8f8bf16_contig_one(
     num_tma_threads = 128
     num_math_threads_per_group = 128
     kwargs = {
+        "GEMM_TYPE": GemmType.GroupedContiguous,
         "NUM_TMA_THREADS": num_tma_threads,
         "NUM_MATH_THREADS_PER_GROUP": num_math_threads_per_group,
+        "N": n,
+        "K": k,
+        "NUM_GROUPS": 1,
+        "BLOCK_M": block_m,
+        "BLOCK_N": block_n,
         "BLOCK_K": block_k,
+        "SWIZZLE_D_MODE": smem_config[1],
+        "BLOCK_N_PADDING": smem_config[2],
+        "NUM_STAGES": num_stages,
+        "NUM_TMA_MULTICAST": tma_multicast_config[0],
+        "IS_TMA_MULTICAST_ON_A": tma_multicast_config[1],
         "NUM_SMS": num_sms,
         "SMEM_SIZE": smem_config[0],
     }
-    _, _ = jit_tuner.compile_and_tune(
-        name="m_grouped_gemm_fp8_fp8_bf16_nt",
-        keys={
-            "N": n,
-            "K": k,
-            "BLOCK_M": block_m,
-            "BLOCK_N": block_n,
-            "SWIZZLE_D_MODE": smem_config[1],
-            "BLOCK_N_PADDING": smem_config[2],
-            "NUM_GROUPS": num_groups,
-            "NUM_STAGES": num_stages,
-            "NUM_TMA_MULTICAST": tma_multicast_config[0],
-            "IS_TMA_MULTICAST_ON_A": tma_multicast_config[1],
-            "GEMM_TYPE": GemmType.GroupedContiguous,
-        },
-        space=(),
-        kwargs=kwargs,
-        runtime_cls=FP8GemmRuntime,
-    )
+
+    code = FP8GemmRuntime.generate(kwargs)
+    _ = build("m_grouped_gemm_fp8_fp8_bf16_nt", code, FP8GemmRuntime, kwargs)
 
 
 def _compile_gemm_nt_f8f8bf16_one(
@@ -228,28 +219,23 @@ def _compile_gemm_nt_f8f8bf16_one(
         "GEMM_TYPE": GemmType.Normal,
         "NUM_TMA_THREADS": num_tma_threads,
         "NUM_MATH_THREADS_PER_GROUP": num_math_threads_per_group,
+        "N": n,
+        "K": k,
         "NUM_GROUPS": 1,
+        "BLOCK_M": block_m,
+        "BLOCK_N": block_n,
         "BLOCK_K": block_k,
+        "SWIZZLE_D_MODE": smem_config[1],
+        "BLOCK_N_PADDING": smem_config[2],
+        "NUM_STAGES": num_stages,
+        "NUM_TMA_MULTICAST": tma_multicast_config[0],
+        "IS_TMA_MULTICAST_ON_A": tma_multicast_config[1],
         "NUM_SMS": num_sms,
         "SMEM_SIZE": smem_config[0],
     }
-    _, _ = jit_tuner.compile_and_tune(
-        name="gemm_fp8_fp8_bf16_nt",
-        keys={
-            "N": n,
-            "K": k,
-            "BLOCK_M": block_m,
-            "BLOCK_N": block_n,
-            "SWIZZLE_D_MODE": smem_config[1],
-            "BLOCK_N_PADDING": smem_config[2],
-            "NUM_STAGES": num_stages,
-            "NUM_TMA_MULTICAST": tma_multicast_config[0],
-            "IS_TMA_MULTICAST_ON_A": tma_multicast_config[1],
-        },
-        space=(),
-        kwargs=kwargs,
-        runtime_cls=FP8GemmRuntime,
-    )
+
+    code = FP8GemmRuntime.generate(kwargs)
+    _ = build("gemm_fp8_fp8_bf16_nt", code, FP8GemmRuntime, kwargs)
 
 
 _KERNEL_HELPER_DICT: Dict[DeepGemmKernelType, DeepGemmKernelHelper] = {
@@ -391,3 +377,16 @@ def _log_jit_build(M: int, N: int, K: int, kernel_type: DeepGemmKernelType):
     RuntimeCache.get = __patched_func
     yield
     RuntimeCache.get = origin_func
+
+
+@contextmanager
+def configure_deep_gemm_num_sms(num_sms):
+    if num_sms is None:
+        yield
+    else:
+        original_num_sms = deep_gemm.get_num_sms()
+        deep_gemm.set_num_sms(num_sms)
+        try:
+            yield
+        finally:
+            deep_gemm.set_num_sms(original_num_sms)