perf: optimize PCG inductor path for FP8 models (redo of #21734) (#23227)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

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

@@ -18,6 +18,7 @@ if TYPE_CHECKING:
 from sglang.srt.layers.quantization.fp8_kernel import (
     fp8_dtype,
     fp8_max,
+    fp8_min,
     is_fp8_fnuz,
     mxfp8_block_scaled_matmul_triton,
     per_token_group_quant_fp8,
@@ -28,6 +29,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     w8a8_block_fp8_matmul_deepgemm,
     w8a8_block_fp8_matmul_triton,
 )
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     ceil_align,
     ceil_div,
@@ -1467,6 +1469,26 @@ def apply_fp8_linear(
         num_token_padding = output_padding
         if cutlass_fp8_supported and weight_scale.numel() == weight.shape[1]:
             num_token_padding = None
+        # For static per-tensor activation scales when using inductor compiler,
+        # use pure PyTorch ops instead of the opaque sgl_kernel quant kernel.
+        # Inductor fuses these with surrounding ops (RMSNorm, residual add),
+        # eliminating a separate kernel launch per linear layer.
+        # weight_scale shape does not matter here -- it is only used in the
+        # GEMM epilogue, not in the activation quant fusion. Only activates when
+        # piecewise_cuda_graph_compiler=inductor; eager PCG and decode both
+        # use the faster custom kernel.
+        if (
+            input_scale is not None
+            and input_scale.numel() == 1
+            and get_global_server_args().piecewise_cuda_graph_compiler == "inductor"
+        ):
+            qinput = (
+                (input_2d * input_scale.reciprocal())
+                .clamp(min=fp8_min, max=fp8_max)
+                .to(fp8_dtype)
+            )
+            x_scale = input_scale
+        else:
             qinput, x_scale = scaled_fp8_quant(
                 input_2d,
                 input_scale,

python/sglang/srt/models/utils.py

@@ -32,6 +32,7 @@ from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
 from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import get_current_device_stream_fast, is_cuda, is_hip
 from sglang.srt.utils.custom_op import register_custom_op
 
@@ -390,6 +391,28 @@ class RotaryPosMixin:
         return torch.from_numpy(np.stack([hpos_ids, wpos_ids], axis=-1))
 
 
+def _reshape_for_qk_norm(x: torch.Tensor, head_dim: int) -> torch.Tensor:
+    """Reshape a (..., H*D) tensor into (..., H, D) ahead of QK RMSNorm.
+
+    On CUDA with the inductor piecewise-cuda-graph compiler, return a
+    stride-preserving view so inductor can fuse this reshape with the
+    subsequent RMSNorm (and any upstream/downstream FP8 quant) into a
+    single triton kernel -- the original motivation of #21734.
+
+    Everywhere else (ROCm, or CUDA with the eager PCG fallback), use the
+    flat 2D reshape that forces a copy when the input is a non-contiguous
+    QKV-split stride-trick view. ROCm's RMSNorm kernels assume contiguous
+    inputs and fault on strided tensors (root cause of the #21734 revert
+    in #23159).
+    """
+    if (
+        _is_cuda
+        and get_global_server_args().piecewise_cuda_graph_compiler == "inductor"
+    ):
+        return x.view(*x.shape[:-1], -1, head_dim)
+    return x.reshape(-1, head_dim)
+
+
 def apply_qk_norm(
     q: torch.Tensor,
     k: torch.Tensor,
@@ -424,6 +447,8 @@ def apply_qk_norm(
         and allow_inplace  # TODO(dark): this can be relaxed if needed
         and (q_eps == k_eps)  # TODO(dark): this can also be relaxed
         and not envs.SGLANG_ENABLE_DETERMINISTIC_INFERENCE.get()
+        and get_global_server_args().piecewise_cuda_graph_compiler
+        != "inductor"  # let inductor fuse QK norm
         and can_use_fused_inplace_qknorm(head_dim, q.dtype)
     ):
         fused_inplace_qknorm(
@@ -439,16 +464,16 @@ def apply_qk_norm(
     if alt_stream is not None and get_is_capture_mode():
         current_stream = get_current_device_stream_fast()
         alt_stream.wait_stream(current_stream)
-        q_by_head = q.reshape(-1, head_dim)
+        q_by_head = _reshape_for_qk_norm(q, head_dim)
         q_by_head = q_norm(q_by_head)
         with torch.cuda.stream(alt_stream):
-            k_by_head = k.reshape(-1, head_dim)
+            k_by_head = _reshape_for_qk_norm(k, head_dim)
             k_by_head = k_norm(k_by_head)
         current_stream.wait_stream(alt_stream)
     else:
-        q_by_head = q.reshape(-1, head_dim)
+        q_by_head = _reshape_for_qk_norm(q, head_dim)
         q_by_head = q_norm(q_by_head)
-        k_by_head = k.reshape(-1, head_dim)
+        k_by_head = _reshape_for_qk_norm(k, head_dim)
         k_by_head = k_norm(k_by_head)
     q = q_by_head.view(q.shape)
     k = k_by_head.view(k.shape)