iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[CI Bug] Revert #42379 to fix CI Multi-Modal Models (Extended Generation 1) (#46070)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-06-18 12:37:39 -07:00
committed by GitHub
parent f6ba720963
commit 225936a1dd
2 changed files with 24 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/libtorch_stable/layernorm_kernels.cu
+7 -6
View File
@@ -81,11 +81,11 @@ __global__ void rms_norm_kernel(
#pragma unroll
for (int j = 0; j < VEC_SIZE; j++) {
float x = static_cast<float>(src1.val[j]);
scalar_t normalized = static_cast<scalar_t>(x * s_variance);
if constexpr (HasWeight) {
dst.val[j] = normalized * src2.val[j];
float w = static_cast<float>(src2.val[j]);
dst.val[j] = static_cast<scalar_t>(x * s_variance * w);
} else {
dst.val[j] = normalized;
dst.val[j] = static_cast<scalar_t>(x * s_variance);
}
}
v_out[i] = dst;
@@ -151,7 +151,8 @@ fused_add_rms_norm_kernel(
#pragma unroll
for (int j = 0; j < width; ++j) {
float x = Converter::convert(res.data[j]);
out.data[j] = Converter::convert(x * s_variance) * w.data[j];
float wf = Converter::convert(w.data[j]);
out.data[j] = Converter::convert(x * s_variance * wf);
}
} else {
#pragma unroll
@@ -198,8 +199,8 @@ fused_add_rms_norm_kernel(
for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
float x = (float)residual[blockIdx.x * hidden_size + idx];
if constexpr (HasWeight) {
input[blockIdx.x * input_stride + idx] =
(scalar_t)(x * s_variance) * weight[idx];
float w = (float)weight[idx];
input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance * w);
} else {
input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance);
}
csrc/libtorch_stable/layernorm_quant_kernels.cu
+17 -6
View File
@@ -66,8 +66,13 @@ __global__ void rms_norm_static_fp8_quant_kernel(
#pragma unroll
for (int j = 0; j < VEC_SIZE; j++) {
float x = static_cast<float>(src1.val[j]);
// Multiply in weight's native dtype to match rms_norm_kernel.
scalar_t out_norm = static_cast<scalar_t>(x * s_variance) * src2.val[j];
float w = static_cast<float>(src2.val[j]);
// Round normalized result through scalar_t to match the precision of the
// unfused composite (rms_norm writes scalar_t, then
// static_scaled_fp8_quant re-loads it as float before FP8 conversion).
// Without this round, the fused path is strictly more accurate and
// disagrees with the composite at exact E4M3 quantization tie boundaries.
scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
out[blockIdx.x * hidden_size + idx * VEC_SIZE + j] =
scaled_fp8_conversion<true, fp8_type>(static_cast<float>(out_norm),
scale_inv);
@@ -137,8 +142,12 @@ fused_add_rms_norm_static_fp8_quant_kernel(
#pragma unroll
for (int i = 0; i < width; ++i) {
float x = Converter::convert(res.data[i]);
// Multiply in weight's native dtype to match fused_add_rms_norm_kernel.
HipT out_norm_h = Converter::convert(x * s_variance) * w.data[i];
float wf = Converter::convert(w.data[i]);
// See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
// to match the unfused composite path at FP8 boundaries. We use the
// backend's hip_type for the intermediate since c10::Half/BFloat16 has
// ambiguous conversions on CUDA and no implicit conversion on ROCm.
HipT out_norm_h = Converter::convert(x * s_variance * wf);
out[id * width + i] = scaled_fp8_conversion<true, fp8_type>(
Converter::convert(out_norm_h), scale_inv);
}
@@ -183,8 +192,10 @@ fused_add_rms_norm_static_fp8_quant_kernel(
for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
float x = (float)residual[blockIdx.x * hidden_size + idx];
// Multiply in weight's native dtype to match fused_add_rms_norm_kernel.
scalar_t out_norm = static_cast<scalar_t>(x * s_variance) * weight[idx];
float w = (float)weight[idx];
// See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
// to match the unfused composite path at FP8 boundaries.
scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
out[blockIdx.x * hidden_size + idx] = scaled_fp8_conversion<true, fp8_type>(
static_cast<float>(out_norm), scale_inv);
}