FP8 DPA future token leakage

先运行最小复现，然后检查 pytorch/attention/dot_product_attention/utils.py、common/fused_attn/fused_attn_fp8.cu 以及 recipe scaling 文件中所引用的 attention 路径。将 MXFP8 和当前 scaling 的行为与 BF16 控制进行比较；当仅更改未来值时因果前缀输出仍保持不变，并且两条路径都有回归覆盖时，即视为完成。

FP8 scaling can leak future tokens into causal SDPA outputs

Summary

Transformer Engine has two current-data FP8 DPA paths in which scale selection can inspect masked future tokens before the causal mask is applied. MXFP8 quantizes V for PV with an E8M0 scale shared by 32 sequence positions, while tensorwise current scaling quantizes the combined Q/K/V tensor with one current amax. In both cases, a future V value can alter the quantized representation of an allowed prefix V value.

Expected behavior

For causal self-attention with identical Q, K, and V through position t, changing V only at positions greater than t should not change outputs through t:

O(Q, K, V_a)[:, :, :t+1, :] == O(Q, K, V_b)[:, :, :t+1, :]
when V_a[:, :, :t+1, :] == V_b[:, :, :t+1, :]

Small nondeterministic kernel variation aside, future masked values should not be an input to the numerical representation of allowed prefix values.

Actual behavior and causes

MXFP8 sequence-block V scaling

For each V block B_b = {32b, ..., 32b+31} and each batch, head, and V feature, TE computes:

D[b,h,d]     = E8M0_round_up(max(abs(V[j,h,d]), j in B_b) / FP8_MAX)
V8[j,h,d]    = cast_fp8(V[j,h,d] / D[b,h,d])
V_hat[j,h,d] = fp32(V8[j,h,d]) * D[b,h,d]

Causal PV then computes:

O[t,h,d] = sum(P[t,h,j] * V_hat[j,h,d], j <= t)

P is not the source of this dependency. In the MXFP8 path, TE nulls the external S and dP quantizer slots, and cuDNN's checked-in online-softmax reference quantizes the unnormalized P tile using fixed, data-independent constants s_scale=16 and inv_s_scale=1/16. The sequence-dependent reduction at issue is V's columnwise E8M0 amax.

A future V[r,h,d], where t < r < 32b+32, can change D[b,h,d], which can change the quantized representation of an allowed V[j,h,d] for j <= t. The future token has no direct PV term because P[t,h,r] = 0, but O[t,h,d] can still change through the shared scale.

Tensorwise current QKV scaling

The current-scaling path has the same causal problem with broader scope. For ordinary dense DPA, TE combines Q, K, and V and applies one current quantizer to the combined tensor:

A         = max(abs(concat(Q, K, V)))
s         = 448 / A
Q8,K8,V8  = round_E4M3(s * concat(Q, K, V))
Q_hat,... = fp32(Q8,K8,V8) / s

Therefore any future Q, K, or V outlier can change the representation of every prefix Q, K, and V value in the same local DPA tensor.

Source evidence

1. TE selects rowwise Q/K but columnwise V for forward MXFP8 attention, so V scales reduce along S: utils.py lines 2693–2727.
2. TE passes V's columnwise payload and columnwise scale factors to the fused backend: fused_attn_fp8.cu lines 1105–1123.
3. cuDNN Frontend fixes the block size at 32 and validates V scales as [b, h, ceil(s_kv/32), d]: scaled_dot_product_flash_attention.h lines 326–413.
4. cuDNN dequantizes V with block size {1, 32} immediately before BMM2/PV: scaled_dot_product_flash_attention.h lines 1026–1046.
5. The causal mask is applied to the QK score tensor inside SDPA, after V has already been quantized externally: scaled_dot_product_flash_attention.h lines 860–884.
6. For non-MX FP8 DPA, TE combines packed or separate Q, K, and V storage and invokes the QKV quantizer once on the combined tensor: utils.py lines 2777–2813.
7. Current scaling computes max_fp8 / amax as a full FP32 value by default and clears the mantissa only when power-of-two scaling is explicitly enabled: recipe_common.cuh lines 14–48 and Float8CurrentScaling defaults at lines 284–304.

Minimal repro

import os

os.environ["NVTE_ALLOW_NONDETERMINISTIC_ALGO"] = "0"
os.environ["NVTE_FLASH_ATTN"] = "0"
os.environ["NVTE_FUSED_ATTN"] = "1"
os.environ["NVTE_UNFUSED_ATTN"] = "0"

import torch
import transformer_engine
from transformer_engine.common.recipe import Float8CurrentScaling, Format, MXFP8BlockScaling
from transformer_engine.pytorch import DotProductAttention, autocast

S, D, T = 128, 128, 7
q = torch.ones((S, 1, 1, D), device="cuda", dtype=torch.bfloat16)
k = torch.ones_like(q)

mx_base_v = torch.zeros_like(q)
mx_base_v[0, 0, 0, 0] = 2.0**-20
mx_same_block_v = mx_base_v.clone()
mx_same_block_v[16, 0, 0, 0] = 1.0
mx_next_block_v = mx_base_v.clone()
mx_next_block_v[32, 0, 0, 0] = 1.0

current_base_v = torch.zeros_like(q)
current_base_v[0, 0, 0, 0] = 2.0**-10
current_future_v = current_base_v.clone()
current_future_v[64, 0, 0, 0] = 1024.0

mx_recipe = MXFP8BlockScaling(
    fp8_format=Format.E4M3,
    fp8_dpa=True,
    fp8_mha=False,
)
current_recipe = Float8CurrentScaling(
    fp8_format=Format.E4M3,
    fp8_dpa=True,
    fp8_mha=False,
)


def run(value, fp8, recipe=None):
    dpa = DotProductAttention(
        1,
        D,
        attention_dropout=0.0,
        qkv_format="sbhd",
        attn_mask_type="causal",
    ).cuda().eval()
    with torch.no_grad(), autocast(enabled=fp8, recipe=recipe):
        return dpa(q, k, value, attn_mask_type="causal").float()

def report(name, reference, changed):
    delta = (reference[: T + 1] - changed[: T + 1]).abs()
    print(
        f"{name:22} max_abs={delta.max().item():.10e} "
        f"changed={delta.count_nonzero().item():2d} "
        f"O[t,0]={reference[T, 0, 0].item():.10e} -> "
        f"{changed[T, 0, 0].item():.10e}"
    )
    return delta.max().item()

mx_base = run(mx_base_v, True, mx_recipe)
mx_same = run(mx_same_block_v, True, mx_recipe)
mx_next = run(mx_next_block_v, True, mx_recipe)
current_base = run(current_base_v, True, current_recipe)
current_future = run(current_future_v, True, current_recipe)
bf16_mx_base = run(mx_base_v, False)
bf16_mx_same = run(mx_same_block_v, False)
bf16_current_base = run(current_base_v, False)
bf16_current_future = run(current_future_v, False)

print(
    f"TE={transformer_engine.__version__} torch={torch.__version__} "
    f"cuDNN={torch.backends.cudnn.version()} GPU={torch.cuda.get_device_name(0)}"
)
same_block_diff = report("MXFP8 same block", mx_base, mx_same)
next_block_diff = report("MXFP8 next block", mx_base, mx_next)
current_diff = report("Current future token", current_base, current_future)
bf16_mx_diff = report("BF16 MX control", bf16_mx_base, bf16_mx_same)
bf16_current_diff = report(
    "BF16 current control", bf16_current_base, bf16_current_future
)

assert next_block_diff == 0.0
assert bf16_mx_diff == 0.0
assert bf16_current_diff == 0.0

failures = []
if same_block_diff != 0.0:
    failures.append("MXFP8 same-block V scale changed the causal prefix")
if current_diff != 0.0:
    failures.append("current QKV tensor scale changed the causal prefix")
assert not failures, "; ".join(failures)

Observed current behavior

Reproduced on B200, Transformer Engine 2.16.0+4220403e, PyTorch 2.13.0a0+8145d630e8.nv26.06, and cuDNN 9.23.0:

TE=2.16.0+4220403e torch=2.13.0a0+8145d630e8.nv26.06 cuDNN=92300 GPU=NVIDIA B200
MXFP8 same block       max_abs=9.5367431641e-07 changed= 8 O[t,0]=1.1920928955e-07 -> 0.0000000000e+00
MXFP8 next block       max_abs=0.0000000000e+00 changed= 0 O[t,0]=1.1920928955e-07 -> 1.1920928955e-07
Current future token   max_abs=9.7656250000e-04 changed= 8 O[t,0]=1.2207031250e-04 -> 0.0000000000e+00
BF16 MX control        max_abs=0.0000000000e+00 changed= 0 O[t,0]=1.1920928955e-07 -> 1.1920928955e-07
BF16 current control   max_abs=0.0000000000e+00 changed= 0 O[t,0]=1.2207031250e-04 -> 1.2207031250e-04
AssertionError: MXFP8 same-block V scale changed the causal prefix; current QKV tensor scale changed the causal prefix