[ROCm][MLA] Fuse MLA q/kv RMSNorm + FP8 per-token quant in the FP8 attention path (#44977)

Signed-off-by: Xavier Aguilar <xavier.aguilarfruto@amd.com>
Signed-off-by: Xavier Aguilar <Xavier.AguilarFruto@amd.com>
This commit is contained in:
xaguilar-amd
2026-07-02 13:00:41 +08:00
committed by GitHub
parent d63c8e9444
commit 09663abde0
4 changed files with 371 additions and 5 deletions
+25 -4
View File
@@ -30,7 +30,7 @@ or just on the low or high end.
| [RMSNorm + Quant](#rmsnorm--quantization-fuse_norm_quant) | `fuse_norm_quant` | RMSNorm (+residual add) → FP8/FP4 quant | O1 (conditional) | 1-4% | No | Always |
| [SiLU+Mul + Quant](#silumul--quantization-fuse_act_quant) | `fuse_act_quant` | SiLU+Mul activation → FP8/FP4 quant | O1 (conditional) | 1-4% | No | Always |
| [RMSNorm + Padding](#rmsnorm--padding-fuse_act_padding) | `fuse_act_padding` | Residual add + RMSNorm → padding | O1 (ROCm/AITER only) | TBD | No | Always |
| [MLA Dual RMSNorm](#mla-dual-rmsnorm-fuse_mla_dual_rms_norm) | `fuse_mla_dual_rms_norm` | Paired Q + KV RMSNorm → single kernel | O1 (ROCm/AITER only) | ~2% | No | Always |
| [MLA Dual RMSNorm](#mla-dual-rmsnorm-fuse_mla_dual_rms_norm) | `fuse_mla_dual_rms_norm` | Paired Q + KV RMSNorm (+ FP8 quant) → 1 kernel | O1 (ROCm/AITER only) | 1-2% | No | Always |
## Support Matrix
@@ -381,11 +381,32 @@ q_normed, kv_normed = fused_mla_dual_rms_norm(
Requires: AMD ROCm with AITER enabled. Enabled by default at optimization level O1 and above
when AITER is available.
**FP8 attention variant (per-token quant).** With a per-token FP8 `q_b_proj`,
only the *q* latent is FP8-quantized while *kv* stays bf16.
`RocmAiterRMSNormQuantFusionPass` first folds the q side into
`rocm_aiter_rmsnorm_fused_dynamic_quant`, leaving kv a plain
`rms_norm` — breaking the symmetric pattern above. The same pass then matches
this asymmetric pair and lowers it to `fused_mla_dual_rms_norm_per_token_quant`.
```text
# Unfused (q norm+quant fused; kv still plain rms_norm):
q_c, kv_lora = split(projected, [q_dim, kv_dim])
kv_c, k_pe = split(kv_lora, [kv_c_dim, k_pe_dim])
q_fp8, q_scale = rocm_aiter_rmsnorm_fused_dynamic_quant(q_c, q_weight, eps, fp8)
kv_normed = rms_norm(kv_c, kv_weight, eps) # bf16
# Fused:
q_c, kv_lora = split(projected, [q_dim, kv_dim])
kv_c, k_pe = split(kv_lora, [kv_c_dim, k_pe_dim])
q_fp8, q_scale, kv_normed = fused_mla_dual_rms_norm_per_token_quant(
q_c, q_weight, kv_c, kv_weight, eps1, eps2)
```
**Code locations.**
- Pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py) (`MLADualRMSNormFusionPass`)
- Custom op: [`vllm/_aiter_ops.py`](https://github.com/vllm-project/vllm/blob/main/vllm/_aiter_ops.py) (`fused_mla_dual_rms_norm`)
- AITER kernel: [`fused_qk_rmsnorm`](https://github.com/ROCm/aiter/pull/2442)
- Pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py) (`MLADualRMSNormFusionPass`, `MLADualRMSPerTokenQuantPattern`)
- Custom op: [`vllm/_aiter_ops.py`](https://github.com/vllm-project/vllm/blob/main/vllm/_aiter_ops.py) (`fused_mla_dual_rms_norm`, `fused_mla_dual_rms_norm_per_token_quant`)
- AITER kernels: [`fused_qk_rmsnorm`](https://github.com/ROCm/aiter/pull/2442), `fused_qk_rmsnorm_per_token_quant`
## See Also
@@ -12,7 +12,10 @@ import torch
import vllm.config
from tests.compile.backend import TestBackend
from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops
from vllm._aiter_ops import (
is_aiter_found_and_supported,
rocm_aiter_ops,
)
from vllm.compilation.passes.utility.noop_elimination import NoOpEliminationPass
from vllm.compilation.passes.utility.post_cleanup import PostCleanupPass
from vllm.config import (
@@ -23,6 +26,7 @@ from vllm.config import (
VllmConfig,
)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.platforms import current_platform
# MLA attention geometry for DeepSeek-V3 / Kimi-K2
Q_DIM = 1536
@@ -30,6 +34,8 @@ KV_C_DIM = 512
K_PE_DIM = 64
EPS = 1e-6
FP8_DTYPE = current_platform.fp8_dtype()
class MLADualRMSNormTestModel(torch.nn.Module):
"""
@@ -146,3 +152,139 @@ def test_fuse_mla_dual_rms_norm(
backend.check_before_ops(model.ops_in_model_before())
backend.check_after_ops(model.ops_in_model_after())
class MLADualRMSNormFp8PerTokenTestModel(torch.nn.Module):
"""
Minimal model reproducing the FP8 MLA attention path with *per-token* quant:
linear -> split([q_dim, kv_dim])
+-- q_c (getitem 0) -> rocm_aiter_rmsnorm_fused_dynamic_quant -> dequant
+-- kv_lora (getitem 1) -> split([kv_c_dim, k_pe_dim])
+-- kv_c (getitem 0) -> rms_norm (bf16)
+-- k_pe
"""
def __init__(
self,
hidden_size: int,
q_dim: int = Q_DIM,
kv_c_dim: int = KV_C_DIM,
k_pe_dim: int = K_PE_DIM,
eps: float = EPS,
):
super().__init__()
self.q_dim = q_dim
self.kv_dim = kv_c_dim + k_pe_dim
self.kv_c_dim = kv_c_dim
self.k_pe_dim = k_pe_dim
self.eps = eps
self.proj = torch.nn.Linear(hidden_size, q_dim + self.kv_dim, bias=False)
self.q_weight = torch.nn.Parameter(torch.ones(q_dim))
self.kv_norm = RMSNorm(kv_c_dim, eps=eps)
def _dequant(self, x_fp8: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
# Per-token: a single (M, 1) scale broadcast across the row.
return (x_fp8.to(torch.float32) * scale).to(torch.bfloat16)
def forward(self, x: torch.Tensor):
# Avoid graph input being a direct arg to a matched pattern node
x = torch.relu(x)
projected = self.proj(x)
q_c, kv_lora = projected.split([self.q_dim, self.kv_dim], dim=-1)
kv_c, k_pe = kv_lora.split([self.kv_c_dim, self.k_pe_dim], dim=-1)
q_fp8, q_scale = torch.ops.vllm.rocm_aiter_rmsnorm_fused_dynamic_quant(
q_c, self.q_weight, self.eps, FP8_DTYPE
)
kv_normed = self.kv_norm(kv_c)
return self._dequant(q_fp8, q_scale), kv_normed, k_pe
def ops_in_model_before(self):
return [
torch.ops.vllm.rocm_aiter_rmsnorm_fused_dynamic_quant.default,
torch.ops.vllm_ir.rms_norm.default,
]
def ops_in_model_after(self):
return [torch.ops.vllm.fused_mla_dual_rms_norm_per_token_quant.default]
@pytest.mark.parametrize("dtype", [torch.bfloat16])
@pytest.mark.parametrize("hidden_size", [7168])
@pytest.mark.skipif(
not is_aiter_found_and_supported(),
reason="Only test on ROCm with AITER installed and supported",
)
def test_fuse_mla_dual_rms_norm_fp8_per_token(
dtype: torch.dtype,
hidden_size: int,
monkeypatch: pytest.MonkeyPatch,
):
torch._dynamo.reset()
vllm_config = VllmConfig(
model_config=ModelConfig(dtype=dtype),
compilation_config=CompilationConfig(
mode=CompilationMode.VLLM_COMPILE,
custom_ops=["+rms_norm"],
pass_config=PassConfig(
fuse_mla_dual_rms_norm=True,
eliminate_noops=True,
),
),
)
with vllm.config.set_current_vllm_config(vllm_config), monkeypatch.context() as m:
from vllm.compilation.passes.fusion.rocm_aiter_fusion import (
MLADualRMSNormFusionPass,
)
torch.set_default_device("cuda")
torch.set_default_dtype(dtype)
torch.manual_seed(42)
m.setenv("VLLM_ROCM_USE_AITER", "1")
rocm_aiter_ops.refresh_env_variables()
fusion_pass = MLADualRMSNormFusionPass(vllm_config)
passes = [
NoOpEliminationPass(vllm_config),
fusion_pass,
PostCleanupPass(vllm_config),
]
backend = TestBackend(*passes)
model = MLADualRMSNormFp8PerTokenTestModel(hidden_size)
x = torch.randn(4, hidden_size)
torch._dynamo.mark_dynamic(x, 0)
with torch.inference_mode():
outputs_unfused = model(x)
model_fused = torch.compile(model, backend=backend)
outputs_fused = model_fused(x)
q_deq_u, kv_normed_u, k_pe_u = outputs_unfused
q_deq_f, kv_normed_f, k_pe_f = outputs_fused
torch.testing.assert_close(k_pe_u, k_pe_f, atol=0, rtol=0)
torch.testing.assert_close(kv_normed_u, kv_normed_f, atol=1e-2, rtol=1e-2)
E4M3_STEP = 0.125
exact_frac = (q_deq_u == q_deq_f).float().mean().item()
assert exact_frac > 0.99, (
f"q: only {exact_frac:.4f} of elements bit-exact; scales likely differ"
)
torch.testing.assert_close(q_deq_u, q_deq_f, atol=1e-2, rtol=E4M3_STEP)
assert fusion_pass.matched_count == 1, (
f"Expected 1 fused pair, got {fusion_pass.matched_count}"
)
backend.check_before_ops(model.ops_in_model_before())
backend.check_after_ops(model.ops_in_model_after())
+68
View File
@@ -1355,6 +1355,63 @@ def _fused_mla_dual_rms_norm_fake(
return (torch.empty_like(x1), torch.empty_like(x2))
def _fused_mla_dual_rms_norm_per_token_quant_impl(
q: torch.Tensor,
q_weight: torch.Tensor,
kv: torch.Tensor,
kv_weight: torch.Tensor,
q_epsilon: float,
kv_epsilon: float,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Fused MLA q/kv RMSNorm (+ FP8 per-token quant on q) via AITER.
Backs the ``fused_mla_dual_rms_norm_per_token_quant`` custom op used by the
MLA FP8 attention fusion when the q latent is quantized *per token* (a single
``(M, 1)`` scale). Only the *q* latent is FP8 quantized (it feeds the
FP8 ``q_b_proj`` GEMM); the *kv* latent is RMS-normed and consumed by attention as bf16.
"""
from aiter.ops.fused_qk_rmsnorm_group_quant import (
fused_qk_rmsnorm_per_token_quant,
)
mq, nq = q.shape
q_out = torch.empty((mq, nq), dtype=FP8_DTYPE, device=q.device)
q_scale = torch.empty((mq, 1), dtype=torch.float32, device=q.device)
kv_normed = torch.empty(kv.shape, dtype=kv.dtype, device=kv.device)
# q -> RMSNorm + FP8 per-token quant (q slot); kv -> RMSNorm only (k slot).
# `split` views are accepted directly (unit inner stride); the kernel
# handles strided inputs, matching the aiter op-test usage.
fused_qk_rmsnorm_per_token_quant(
q_out_quantized=q_out,
q_out_scale=q_scale,
q=q,
q_weight=q_weight,
q_epsilon=q_epsilon,
k_out=kv_normed,
k=kv,
k_weight=kv_weight,
k_epsilon=kv_epsilon,
gemma_norm=False,
)
return q_out, q_scale, kv_normed
def _fused_mla_dual_rms_norm_per_token_quant_fake(
q: torch.Tensor,
q_weight: torch.Tensor,
kv: torch.Tensor,
kv_weight: torch.Tensor,
q_epsilon: float,
kv_epsilon: float,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
mq, nq = q.shape
q_out = torch.empty((mq, nq), dtype=FP8_DTYPE, device=q.device)
q_scale = torch.empty((mq, 1), dtype=torch.float32, device=q.device)
kv_normed = torch.empty(kv.shape, dtype=kv.dtype, device=kv.device)
return q_out, q_scale, kv_normed
def _rocm_aiter_gemm_a8wfp4_impl(
x: torch.Tensor,
w: torch.Tensor,
@@ -2046,6 +2103,13 @@ class rocm_aiter_ops:
fake_impl=_fused_mla_dual_rms_norm_fake,
)
direct_register_custom_op(
op_name="fused_mla_dual_rms_norm_per_token_quant",
op_func=_fused_mla_dual_rms_norm_per_token_quant_impl,
mutates_args=[],
fake_impl=_fused_mla_dual_rms_norm_per_token_quant_fake,
)
_OPS_REGISTERED = True
@staticmethod
@@ -2120,6 +2184,10 @@ class rocm_aiter_ops:
def get_fused_mla_dual_rms_norm_op() -> OpOverload:
return torch.ops.vllm.fused_mla_dual_rms_norm.default
@staticmethod
def get_fused_mla_dual_rms_norm_per_token_quant_op() -> OpOverload:
return torch.ops.vllm.fused_mla_dual_rms_norm_per_token_quant.default
@staticmethod
def w8a8_gemm(
A: torch.Tensor,
@@ -922,11 +922,145 @@ class MLADualRMSNormPattern(
return _replacement
class MLADualRMSPerTokenQuantPattern(
VllmPatternReplacement[
...,
tuple[
torch.Tensor,
torch.Tensor,
torch.Tensor,
torch.Tensor,
],
]
):
"""
Fuse the MLA FP8 attention path -- q-latent RMSNorm + FP8 *per-token* quant
plus kv-latent RMSNorm -- into AITER's ``fused_qk_rmsnorm_per_token_quant``.
With a per-token FP8 ``q_b_proj`` (Quark / ModelOpt), the earlier
``RocmAiterRMSNormQuantFusionPass`` folds the q side into
``rocm_aiter_rmsnorm_fused_dynamic_quant`` and leaves the kv side a plain
``vllm_ir.rms_norm``. This pattern matches that asymmetric pair::
gemm -> split_with_sizes([q_dim, kv_dim])
+-- q_c -> rocm_aiter_rmsnorm_fused_dynamic_quant -> (q_fp8, q_scale)
+-- kv_lora -> split_with_sizes([kv_c_dim, k_pe_dim])
+-- kv_c -> vllm_ir.rms_norm -> kv_normed (bf16)
+-- k_pe
"""
DYNAMIC_QUANT_OP = rocm_aiter_ops.get_rmsnorm_fused_dynamic_quant_op()
FUSED_OP = rocm_aiter_ops.get_fused_mla_dual_rms_norm_per_token_quant_op()
def __init__(self, epsilon: float) -> None:
self._epsilon = epsilon
def get_inputs(self) -> list[torch.Tensor]:
q_dim, kv_c_dim, k_pe_dim = 256, 128, 64
return [
self.empty_bf16(5, q_dim + kv_c_dim + k_pe_dim),
self.empty_bf16(q_dim),
self.empty_bf16(kv_c_dim),
]
@property
def pattern(
self,
) -> Callable[
...,
tuple[
torch.Tensor,
torch.Tensor,
torch.Tensor,
torch.Tensor,
],
]:
eps = self._epsilon
dynamic_quant_op = self.DYNAMIC_QUANT_OP
def _pattern(
projected: torch.Tensor,
q_weight: torch.Tensor,
kv_weight: torch.Tensor,
) -> tuple[
torch.Tensor,
torch.Tensor,
torch.Tensor,
torch.Tensor,
]:
q_dim = q_weight.shape[0]
kv_dim = projected.shape[-1] - q_dim
kv_c_dim = kv_weight.shape[0]
k_pe_dim = kv_dim - kv_c_dim
q_c, kv_lora = projected.split([q_dim, kv_dim], dim=-1)
kv_c, k_pe = kv_lora.split([kv_c_dim, k_pe_dim], dim=-1)
q_quant = dynamic_quant_op(
x=q_c,
weight=q_weight,
epsilon=eps,
quant_dtype=FP8_DTYPE,
)
kv_normed = vllm.ir.ops.rms_norm(kv_c, kv_weight, eps)
return q_quant[0], q_quant[1], kv_normed, k_pe
return _pattern
@property
def replacement(
self,
) -> Callable[
...,
tuple[
torch.Tensor,
torch.Tensor,
torch.Tensor,
torch.Tensor,
],
]:
eps = self._epsilon
fused_op = self.FUSED_OP
def _replacement(
projected: torch.Tensor,
q_weight: torch.Tensor,
kv_weight: torch.Tensor,
) -> tuple[
torch.Tensor,
torch.Tensor,
torch.Tensor,
torch.Tensor,
]:
q_dim = q_weight.shape[0]
kv_dim = projected.shape[-1] - q_dim
kv_c_dim = kv_weight.shape[0]
k_pe_dim = kv_dim - kv_c_dim
q_c, kv_lora = projected.split([q_dim, kv_dim], dim=-1)
kv_c, k_pe = kv_lora.split([kv_c_dim, k_pe_dim], dim=-1)
at = fused_op(
q_c,
q_weight,
kv_c,
kv_weight,
eps,
eps,
)
# q_fp8, q_scale, kv_normed, k_pe
return at[0], at[1], at[2], k_pe
return _replacement
class MLADualRMSNormFusionPass(VllmFusionPatternMatcherPass):
"""
Post-grad PatternMatcher pass that fuses paired q / kv RMS norms in
MLA attention into ``fused_mla_dual_rms_norm`` backed by aiter's
``fused_qk_rmsnorm`` HIP kernel.
The FP8 attention path is also handled via
:class:`MLADualRMSPerTokenQuantPattern`, which fuses the q-latent RMSNorm +
FP8 per-token quant together with the kv-latent RMSNorm into
``fused_mla_dual_rms_norm_per_token_quant`` backed by aiter's
``fused_qk_rmsnorm_per_token_quant`` HIP kernel.
"""
def __init__(self, config: VllmConfig) -> None:
@@ -934,3 +1068,4 @@ class MLADualRMSNormFusionPass(VllmFusionPatternMatcherPass):
for epsilon in [1e-5, 1e-6]:
self.register(MLADualRMSNormPattern(epsilon))
self.register(MLADualRMSPerTokenQuantPattern(epsilon))