AITER backend extension and workload optimizations (#6838)

Co-authored-by: wunhuang <wunhuang@amd.com>
Co-authored-by: Hubert Lu <Hubert.Lu@amd.com>

.github/workflows/pr-test-amd.yml

@@ -72,7 +72,7 @@ jobs:
       - name: Evaluate accuracy (TP=2)
         timeout-minutes: 30
         run: |
-          bash scripts/amd_ci_exec.sh python3 test_moe_eval_accuracy_large.py
+          bash scripts/amd_ci_exec.sh -e SGLANG_USE_AITER=0 python3 test_moe_eval_accuracy_large.py
 
   mla-test-1-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&

docs/references/environment_variables.md

@@ -53,7 +53,7 @@ SGLang supports various environment variables that can be used to configure its
 
 | Environment Variable | Description | Default Value |
 | --- | --- | --- |
-| `SGLANG_AITER_MOE` | Use AITER MOE implementation | `false` |
+| `SGLANG_USE_AITER` | Use AITER optimize implementation | `false` |
 | `SGLANG_INT4_WEIGHT` | Enable INT4 weight quantization | `false` |
 | `SGLANG_MOE_PADDING` | Enable MoE padding (sets padding size to 128 if value is `1`, often set to `1` in Docker builds) | `0` |
 | `SGLANG_FORCE_FP8_MARLIN` | Force using FP8 MARLIN kernels even if other FP8 kernels are available | `false` |

python/sglang/srt/layers/layernorm.py

@@ -20,10 +20,11 @@ import torch
 import torch.nn as nn
 
 from sglang.srt.custom_op import CustomOp
-from sglang.srt.utils import is_cuda, is_hip
+from sglang.srt.utils import get_bool_env_var, is_cuda, is_hip
 
 _is_cuda = is_cuda()
 _is_hip = is_hip()
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _is_cuda:
     from sgl_kernel import (
@@ -33,7 +34,10 @@ if _is_cuda:
         rmsnorm,
     )
 
-if _is_hip:
+if _use_aiter:
+    from aiter import rmsnorm2d_fwd as rms_norm
+    from aiter import rmsnorm2d_fwd_with_add as fused_add_rms_norm
+elif _is_hip:
     from vllm._custom_ops import fused_add_rms_norm, rms_norm
 
 logger = logging.getLogger(__name__)
@@ -48,6 +52,8 @@ class RMSNorm(CustomOp):
         super().__init__()
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
+        if _use_aiter:
+            self._forward_method = self.forward_aiter
 
     def forward_cuda(
         self,
@@ -60,6 +66,25 @@ class RMSNorm(CustomOp):
         out = rmsnorm(x, self.weight.data, self.variance_epsilon)
         return out
 
+    def forward_aiter(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if residual is not None:
+            residual_out = torch.empty_like(x)
+            output = torch.empty_like(x)
+            fused_add_rms_norm(
+                output,
+                x,
+                residual,
+                residual_out,
+                self.weight.data,
+                self.variance_epsilon,
+            )
+            return output, residual_out
+        return rms_norm(x, self.weight.data, self.variance_epsilon)
+
     def forward_hip(
         self,
         x: torch.Tensor,

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -1332,7 +1332,7 @@ def fused_experts_impl(
     if (
         not (use_fp8_w8a8 or use_int8_w8a8)
         or block_shape is not None
-        or (_is_hip and get_bool_env_var("SGLANG_AITER_MOE"))
+        or (_is_hip and get_bool_env_var("SGLANG_USE_AITER"))
     ):
         padded_size = 0
 

python/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,
@@ -188,7 +189,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             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 (

python/sglang/srt/layers/quantization/fp8.py

@@ -77,8 +77,8 @@ _is_cuda = is_cuda()
 
 _is_fp8_fnuz = is_fp8_fnuz()
 
-use_hip_int4 = get_bool_env_var("SGLANG_INT4_WEIGHT")
-use_aiter_moe = get_bool_env_var("SGLANG_AITER_MOE")
+_use_hip_int4 = get_bool_env_var("SGLANG_INT4_WEIGHT")
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _is_hip:
     from aiter import ActivationType, QuantType
@@ -487,7 +487,7 @@ class Fp8MoEMethod:
         from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
 
         if self.quant_config.is_checkpoint_fp8_serialized:
-            params_dtype = torch.uint32 if use_hip_int4 else torch.float8_e4m3fn
+            params_dtype = torch.uint32 if _use_hip_int4 else torch.float8_e4m3fn
         tp_size = get_tensor_model_parallel_world_size()
         if self.block_quant:
             block_n, block_k = (
@@ -512,7 +512,7 @@ class Fp8MoEMethod:
                     )
 
         # WEIGHTS
-        if _is_hip and use_hip_int4:
+        if _is_hip and _use_hip_int4:
             # INT4 MoE weight - INT32 packed
             w13_weight = torch.nn.Parameter(
                 torch.empty(
@@ -641,7 +641,7 @@ class Fp8MoEMethod:
             layer.register_parameter("w13_weight_scale", w13_weight_scale)
             layer.register_parameter("w2_weight_scale", w2_weight_scale)
 
-            if _is_hip:  # and use_aiter_moe: TODO: add check back after triton kernel
+            if _is_hip:  # _use_aiter: TODO: add check back after triton kernel
                 # ROCm - using column scaling, duplicate scaling numbers in case per tensor scaling
                 w13_weight_scale1 = torch.nn.Parameter(
                     torch.ones(num_experts, 2 * intermediate_size, dtype=torch.float32),
@@ -668,7 +668,7 @@ class Fp8MoEMethod:
             set_weight_attrs(w13_weight_scale, extra_weight_attrs)
             set_weight_attrs(w2_weight_scale, extra_weight_attrs)
 
-            if _is_hip and use_hip_int4:
+            if _is_hip and _use_hip_int4:
                 extra_weight_attrs.update(
                     {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
                 )
@@ -700,7 +700,7 @@ class Fp8MoEMethod:
             layer.w2_input_scale = None
 
     def process_weights_after_loading(self, layer: Module) -> None:
-        if _is_hip and use_hip_int4:
+        if _is_hip and _use_hip_int4:
             self.process_weights_hip_int4(layer)
             return
 
@@ -731,7 +731,7 @@ class Fp8MoEMethod:
                 )
                 layer.w2_input_scale = None
 
-            if _is_hip and use_aiter_moe:
+            if _use_aiter:
                 # Pre-shuffle weights
                 layer.w13_weight.data = shuffle_weight(
                     layer.w13_weight.contiguous(), (16, 16)
@@ -853,7 +853,7 @@ class Fp8MoEMethod:
             return
 
     def process_weights_hip_int4(self, layer: Module):
-        # TODO: and use_aiter_moe: add after triton kernel added
+        # TODO: _use_aiter: add after triton kernel added
         # INT4-FP8 (INT4 MoE Weight, FP8 Compute)
         # Weight Permutation
         layer.w13_weight = torch.nn.Parameter(
@@ -900,7 +900,7 @@ class Fp8MoEMethod:
             padding_size,  # Avoid circular import
         )
 
-        if use_aiter_moe:
+        if _use_aiter:
             layer.w13_weight = torch.nn.Parameter(
                 shuffle_weight(layer.w13_weight.data, (16, 16)),
                 requires_grad=False,
@@ -911,7 +911,7 @@ class Fp8MoEMethod:
                 requires_grad=False,
             )
             torch.cuda.empty_cache()
-            # ROCm (use_aiter_moe): using column-wise scaling
+            # ROCm (_use_aiter): using column-wise scaling
             layer.w13_weight_scale1 *= layer.w13_weight_scale.unsqueeze(-1)
             layer.w2_weight_scale1 *= layer.w2_weight_scale.unsqueeze(-1)
         elif get_bool_env_var("SGLANG_MOE_PADDING"):
@@ -1041,8 +1041,8 @@ class Fp8MoEMethod:
         activation: str = "silu",
         no_combine: bool = False,
     ) -> Optional[torch.Tensor]:
-        if use_hip_int4:
-            # TODO: add triton kernel and add check use_aiter_moe
+        if _use_hip_int4:
+            # TODO: add triton kernel and add check _use_aiter
             assert not no_combine, f"{no_combine=} is not supported."
             return ck_moe_2stages(
                 x,
@@ -1058,13 +1058,13 @@ class Fp8MoEMethod:
                 ),
             )
 
-        if use_aiter_moe:
+        if _use_aiter:
             assert not no_combine, f"{no_combine=} is not supported."
             if self.block_quant:
-                # TODO(use_aiter_moe): FP8 block_quant only supports 'silu' for the time-being.
+                # TODO(_use_aiter): FP8 block_quant only supports 'silu' for the time-being.
                 assert (
                     activation == "silu"
-                ), f"use_aiter_moe: FP8 bloack_quant {activation=} will be supported later, unset use_aiter_moe"
+                ), f"_use_aiter: FP8 bloack_quant {activation=} will be supported later, unset _use_aiter"
                 return asm_moe(
                     x,
                     layer.w13_weight,

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -38,11 +38,10 @@ _is_hip = is_hip()
 _is_cuda = is_cuda()
 _is_fp8_fnuz = is_fp8_fnuz()
 
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
-use_aiter_moe = get_bool_env_var("SGLANG_AITER_MOE")
-
-if _is_hip and use_aiter_moe:
-    from aiter import gemm_a8w8_blockscale
+if _use_aiter:
+    from aiter import gemm_a8w8_blockscale_CK
 
 if _is_cuda:
     from sgl_kernel import fp8_blockwise_scaled_mm, fp8_scaled_mm
@@ -141,7 +140,7 @@ def dispatch_w8a8_block_fp8_linear() -> Callable:
         return flashinfer_gemm_w8a8_block_fp8_linear
     elif CUTLASS_BLOCK_FP8_SUPPORTED:
         return cutlass_w8a8_block_fp8_linear_with_fallback
-    elif _is_hip and use_aiter_moe:
+    elif _use_aiter:
         return aiter_w8a8_block_fp8_linear
     elif _ENABLE_JIT_DEEPGEMM:
         return deepgemm_w8a8_block_fp8_linear_with_fallback
@@ -268,12 +267,9 @@ def aiter_w8a8_block_fp8_linear(
     q_input, x_scale = per_token_group_quant_fp8(
         input_2d, block_size[1], column_major_scales=False
     )
-    output = torch.zeros(
-        [q_input.shape[0], weight.shape[0]],
-        dtype=input_2d.dtype,
-        device=q_input.device,
+    output = gemm_a8w8_blockscale_CK(
+        q_input, weight, x_scale, weight_scale, dtype=input.dtype
     )
-    gemm_a8w8_blockscale(q_input, weight, x_scale, weight_scale, output)
 
     if bias is not None:
         output += bias

python/sglang/srt/model_executor/model_runner.py

@@ -355,6 +355,15 @@ class ModelRunner:
                 # MLA architecture
                 if is_hopper_with_cuda_12_3():
                     server_args.attention_backend = "fa3"
+                elif _is_hip:
+                    head_num = self.model_config.get_num_kv_heads(self.tp_size)
+                    # TODO current aiter only support head number 16 or 128 head number
+                    if (
+                        head_num == 128 or head_num == 16
+                    ) and self.spec_algorithm.is_none():
+                        server_args.attention_backend = "aiter"
+                    else:
+                        server_args.attention_backend = "triton"
                 else:
                     server_args.attention_backend = "triton"
             logger.info(
@@ -363,6 +372,7 @@ class ModelRunner:
         elif self.use_mla_backend:
             if server_args.device != "cpu":
                 if server_args.attention_backend in [
+                    "aiter",
                     "flashinfer",
                     "fa3",
                     "triton",

python/sglang/srt/models/deepseek_v2.py

@@ -105,6 +105,7 @@ from sglang.srt.utils import (
 _is_hip = is_hip()
 _is_cuda = is_cuda()
 _is_fp8_fnuz = is_fp8_fnuz()
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _is_cuda:
     from sgl_kernel import awq_dequantize, bmm_fp8, merge_state_v2
@@ -120,6 +121,9 @@ if _is_hip:
         decode_attention_fwd_grouped_rope,
     )
 
+if _use_aiter:
+    from aiter.rotary_embedding import get_rope
+
 logger = logging.getLogger(__name__)
 
 
@@ -697,6 +701,7 @@ class DeepseekV2AttentionMLA(nn.Module):
         )
 
         self.alt_stream = alt_stream
+        self.attn_mha.kv_b_proj = None
 
         self.w_kc = None
         self.w_vc = None
@@ -766,6 +771,15 @@ class DeepseekV2AttentionMLA(nn.Module):
                 return AttnForwardMethod.MHA_CHUNKED_KV
             else:
                 return _dispatch_mla_subtype()
+        elif self.attention_backend == "aiter":
+            if (
+                forward_batch.forward_mode.is_extend()
+                and not forward_batch.forward_mode.is_target_verify()
+                and not forward_batch.forward_mode.is_draft_extend()
+            ):
+                return AttnForwardMethod.MHA
+            else:
+                return AttnForwardMethod.MLA
         else:
             # Triton: Use normal computation for prefill and use weight absorption for extend/decode
             if (
@@ -813,6 +827,9 @@ class DeepseekV2AttentionMLA(nn.Module):
         forward_batch: ForwardBatch,
         zero_allocator: BumpAllocator,
     ):
+        if self.attn_mha.kv_b_proj is None:
+            self.attn_mha.kv_b_proj = self.kv_b_proj
+
         if hidden_states.shape[0] == 0:
             assert (
                 not self.o_proj.reduce_results

scripts/amd_ci_exec.sh

 #!/bin/bash
 set -euo pipefail
 
-# Default working directory
 WORKDIR="/sglang-checkout/test/srt"
-ENV_ARGS=(
-  -e SGLANG_AMD_CI=1
-  -e SGLANG_IS_IN_CI=1
-  -e SGLANG_AITER_MOE=1
+declare -A ENV_MAP=(
+  [SGLANG_AMD_CI]=1
+  [SGLANG_IS_IN_CI]=1
+  [SGLANG_USE_AITER]=1
 )
 
-# Parse optional -w/--workdir and -e ENV=VAL flags
+# Parse -w/--workdir and -e ENV=VAL
 while [[ $# -gt 0 ]]; do
   case "$1" in
     -w|--workdir)
@@ -17,7 +16,8 @@ while [[ $# -gt 0 ]]; do
       shift 2
       ;;
     -e)
-      ENV_ARGS+=("-e" "$2")
+      IFS="=" read -r key val <<< "$2"
+      ENV_MAP["$key"]="$val"
       shift 2
       ;;
     --)
@@ -30,6 +30,12 @@ while [[ $# -gt 0 ]]; do
   esac
 done
 
+# Build final ENV_ARGS
+ENV_ARGS=()
+for key in "${!ENV_MAP[@]}"; do
+  ENV_ARGS+=("-e" "$key=${ENV_MAP[$key]}")
+done
+
 # Run docker exec
 docker exec \
   -w "$WORKDIR" \

test/srt/test_nightly_gsm8k_eval_amd.py

@@ -171,7 +171,7 @@ class TestNightlyGsm8KEval(unittest.TestCase):
                     os.environ["HF_HUB_DISABLE_XET"] = (
                         "1" if model in DISABLE_HF_XET_MODELS else "0"
                     )
-                    os.environ["SGLANG_AITER_MOE"] = (
+                    os.environ["SGLANG_USE_AITER"] = (
                         "0" if model in TRITON_MOE_MODELS else "1"
                     )