[AMD][DSv4] Fuse inverse-RoPE into the fp8 wo_a quant (stacked on #37423) (#37658)

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

python/sglang/srt/environ.py

@@ -1402,6 +1402,10 @@ class Envs:
 
     # cache, GEMM, and distributed
     SGLANG_OPT_FP8_WO_A_GEMM = EnvBool(True)
+    # ROCm gfx950: fuse inverse-RoPE into the wo_a mxfp8 quant (aiter
+    # inverse_rope_group_quant) instead of a separate fused_rope_inplace + Triton
+    # quant. Off by default; requires SGLANG_OPT_FP8_WO_A_GEMM and the aiter op.
+    SGLANG_OPT_FP8_WO_A_FUSED_INVROPE = EnvBool(False)
     # Route the decode wo_a bf16 batched matmul off rocBLAS/Tensile onto aiter's
     # tuned batched_gemm_bf16 (gfx95). Off by default; see deepseek_v4.py
     # _apply_wo_a_bf16_matmul.

python/sglang/srt/models/deepseek_common/amd/deepseek_v4_wo_a_fp8.py

@@ -65,12 +65,38 @@ if _is_hip and _is_gfx95_supported:
             err,
         )
 
+# Fused inverse-RoPE + per-token-group mxfp8 quant in a single aiter kernel.
+# Optional enhancement over the two-kernel path (fused_rope_inplace + the Triton
+# quantizer above): it removes one launch and one HBM round-trip of the [T,H,Dh]
+# attention output. Resolved once at import; None leaves the two-kernel path in
+# place, so this is purely additive.
+_inverse_rope_group_quant = None
+if _batched_gemm_a8w8_mxscale is not None:
+    try:
+        from aiter.ops.inverse_rope_group_quant import (
+            inverse_rope_group_quant as _inverse_rope_group_quant,
+        )
+    except Exception as err:  # pragma: no cover - env-dependent
+        logger.warning(
+            "aiter inverse_rope_group_quant import failed; the DSV4 wo_a fp8 "
+            "fused inverse-RoPE path is unavailable, using the two-kernel "
+            "path: %s",
+            err,
+        )
+
 
 def is_wo_a_fp8_mxscale_supported() -> bool:
     """True when the ROCm fp8 ``wo_a`` path can run on this build/arch."""
     return _batched_gemm_a8w8_mxscale is not None
 
 
+def is_wo_a_fp8_fused_invrope_supported() -> bool:
+    """True when the fused inverse-RoPE + quant ``wo_a`` path can run."""
+    return (
+        _batched_gemm_a8w8_mxscale is not None and _inverse_rope_group_quant is not None
+    )
+
+
 @triton.jit
 def _wo_a_quant_mxfp8_kernel(
     x_ptr,
@@ -160,6 +186,50 @@ def apply_wo_a_fp8_mxscale(
     )
 
 
+def apply_wo_a_fp8_mxscale_fused_invrope(
+    o: torch.Tensor,
+    positions: torch.Tensor,
+    cos_cache: torch.Tensor,
+    sin_cache: torch.Tensor,
+    num_groups: int,
+    weight: torch.Tensor,
+    weight_scale: torch.Tensor,
+) -> torch.Tensor:
+    """fp8 ``wo_a`` with a fused inverse-RoPE + group-quant front end.
+
+    Identical result to ``apply_wo_a_fp8_mxscale`` preceded by the standalone
+    inverse RoPE, but ``inverse_rope_group_quant`` does the inverse rotation and
+    the [T,G,D] mxfp8 quant in one kernel, so the caller must NOT run the
+    separate ``fused_rope_inplace`` first.
+
+    Args:
+        o: ``[T, H, head_dim]`` bf16 attention output, still RoPE'd (the kernel
+            applies the inverse rotation to the trailing rope dims internally).
+        positions: ``[T]`` absolute positions into the cos/sin tables.
+        cos_cache/sin_cache: ``[max_pos, rope_dim // 2]`` bf16 tables.
+        num_groups: local o-groups ``G``.
+        weight: ``[G, R, D]`` fp8 ``wo_a`` weight.
+        weight_scale: ``[G, R/128, D/128]`` uint8 e8m0 (from load-time convert).
+
+    Returns bf16 ``[T, G, R]``.
+    """
+    # Default scale layout is the unshuffled row layout on both the current
+    # (``scale_shuffle=False``) and older (``scale_layout="row"``) aiter
+    # signatures, which is exactly what ``batched_gemm_a8w8_mxscale`` consumes,
+    # so we omit the layout kwarg to stay compatible across aiter versions.
+    o_fp8, o_scale = _inverse_rope_group_quant(
+        o,
+        positions,
+        cos_cache,
+        sin_cache,
+        num_groups=num_groups,
+        quant_group_size=WO_A_MXFP8_GROUP_SIZE,
+    )
+    return _batched_gemm_a8w8_mxscale(
+        o_fp8, weight, o_scale, weight_scale, dtype=torch.bfloat16
+    )
+
+
 def _is_power_of_two(scale: torch.Tensor) -> bool:
     """True when every fp32 scale is an exact power of two (zero mantissa)."""
     bits = scale.detach().float().contiguous().view(torch.int32)

python/sglang/srt/models/deepseek_v4.py

@@ -392,10 +392,13 @@ _wo_a_aiter_batched_gemm_disabled = False
 # the kernel availability and the weight-scale converter resolve once at import;
 # ``None`` here means the platform keeps the bf16 absorb GEMM.
 _wo_a_fp8_mxscale = None
+_wo_a_fp8_mxscale_fused_invrope = None
 _wo_a_weight_scale_to_e8m0 = None
 if _is_hip:
     from sglang.srt.models.deepseek_common.amd.deepseek_v4_wo_a_fp8 import (
         apply_wo_a_fp8_mxscale,
+        apply_wo_a_fp8_mxscale_fused_invrope,
+        is_wo_a_fp8_fused_invrope_supported,
         is_wo_a_fp8_mxscale_supported,
         wo_a_weight_scale_to_e8m0,
     )
@@ -403,6 +406,13 @@ if _is_hip:
     if is_wo_a_fp8_mxscale_supported():
         _wo_a_fp8_mxscale = apply_wo_a_fp8_mxscale
         _wo_a_weight_scale_to_e8m0 = wo_a_weight_scale_to_e8m0
+        # Opt-in fused inverse-RoPE + quant front end (env-gated for A/B). Only
+        # bind it when both the flatmm and the fused aiter op are available.
+        if (
+            envs.SGLANG_OPT_FP8_WO_A_FUSED_INVROPE.get()
+            and is_wo_a_fp8_fused_invrope_supported()
+        ):
+            _wo_a_fp8_mxscale_fused_invrope = apply_wo_a_fp8_mxscale_fused_invrope
 
 
 def _apply_wo_a_bf16_matmul(
@@ -1757,6 +1767,27 @@ class MQALayer(MqaAttentionBase):
                     save_kv_cache=save_kv_cache,
                 )
             o = o[:, tp_slice, :]
+        if (
+            _FP8_WO_A_GEMM
+            and _wo_a_fp8_mxscale_fused_invrope is not None
+            and not _is_npu
+        ):
+            # ROCm gfx950 fused path: inverse-RoPE + per-token-group mxfp8 quant
+            # in one aiter kernel on the pre-view [T,H,Dh] output, then the a8w8
+            # mxscale absorb GEMM. Replaces the standalone inverse RoPE, the
+            # [T,G,D] view, and the quant inside the two-kernel fp8 path below.
+            G = self.n_local_groups
+            cos_c, sin_c = _freqs_cis_to_cos_sin(self.freqs_cis, o.dtype, o.device)
+            o = _wo_a_fp8_mxscale_fused_invrope(
+                o,
+                positions,
+                cos_c,
+                sin_c,
+                G,
+                self.wo_a.weight.view(G, self.o_lora_rank, -1),
+                self.wo_a.weight_scale_inv.data,
+            )
+        else:
             if _is_npu:
                 cos4, sin4 = self._get_npu_rope_position_cache(
                     positions, o.dtype, inverse=True
@@ -1780,9 +1811,9 @@ class MQALayer(MqaAttentionBase):
             o = o.view(o.shape[0], self.n_local_groups, -1)
 
             if _FP8_WO_A_GEMM and _wo_a_fp8_mxscale is not None:
-            # ROCm gfx950: same fp8 absorb GEMM as the DeepGEMM path below, but
+                # ROCm gfx950: same fp8 absorb GEMM as the DeepGEMM path below,
-            # through aiter's e8m0 block-scale batched GEMM. The activation is
+                # but through aiter's e8m0 block-scale batched GEMM. The
-            # quantized per token-group inside the helper.
+                # activation is quantized per token-group inside the helper.
                 T, G, D = o.shape
                 o = _wo_a_fp8_mxscale(
                     o,
@@ -1814,7 +1845,10 @@ class MQALayer(MqaAttentionBase):
                 deep_gemm.fp8_einsum(
                     "bhr,hdr->bhd",
                     (o_fp8, o_s),
-                (self.wo_a.weight.view(G, R, D), self.wo_a.weight_scale_inv.data),
+                    (
+                        self.wo_a.weight.view(G, R, D),
+                        self.wo_a.weight_scale_inv.data,
+                    ),
                     output,
                     recipe=recipe,
                 )