[Bugfix]fix output Nan/Inf in marlin if dtype=float16 (#33972)

Signed-off-by: IriKa Qiu <qiujie.jq@gmail.com>
This commit is contained in:
IriKa
2026-03-27 16:36:08 -07:00
committed by GitHub
parent b69bf2f0b1
commit 148a5c1226
8 changed files with 83 additions and 55 deletions
-3
View File
@@ -189,10 +189,7 @@ __device__ __forceinline__ void cp_async_wait<0>() {
}
__device__ __forceinline__ float clip(float v, float mmin, float mmax) {
#if __CUDACC_VER_MAJOR__ >= 11 && __CUDA_ARCH__ >= 800
return fminf(mmax, fmaxf(v, mmin));
#else
#endif
}
__device__ __forceinline__ __nv_bfloat16 clip(__nv_bfloat16 v,
+1 -1
View File
@@ -13,7 +13,7 @@
const int4 *__restrict__ b_bias_ptr, \
const float *__restrict__ a_scales_ptr, \
const int4 *__restrict__ scales_ptr, \
const uint16_t *__restrict__ global_scale_ptr, \
const float *__restrict__ global_scale_ptr, \
const int4 *__restrict__ zp_ptr, const int *__restrict__ g_idx, \
int num_groups, int prob_m, int prob_n, int prob_k, int lda, int *locks, \
bool has_bias, bool use_atomic_add, bool use_fp32_reduce, \
+5 -5
View File
@@ -57,7 +57,7 @@ torch::Tensor marlin_gemm(
int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
bool is_zp_float) {
TORCH_CHECK_NOT_IMPLEMENTED(false,
"marlin_gemm(..) requires CUDA_ARCH >= 8.0");
"marlin_gemm(..) requires CUDA_ARCH >= 7.5");
return torch::empty({1, 1});
}
@@ -356,7 +356,7 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
const int4* bias_ptr = (const int4*)b_bias;
const float* a_s_ptr = (const float*)a_s;
const int4* b_s_ptr = (const int4*)b_s;
const uint16_t* g_s_ptr = (const uint16_t*)g_s;
const float* g_s_ptr = (const float*)g_s;
const int4* zp_ptr = (const int4*)zp;
const int* g_idx_ptr = (const int*)g_idx;
@@ -751,7 +751,7 @@ torch::Tensor marlin_gemm(
TORCH_CHECK(b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn,
"global_scale can only be used for nvfp4 format.");
} else {
global_scale = torch::empty({0}, options);
global_scale = torch::empty({0}, options_fp32);
TORCH_CHECK(!(b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn),
"the global_scale parameter must be passed for nvfp4 format.");
}
@@ -832,8 +832,8 @@ torch::Tensor marlin_gemm(
TORCH_CHECK(a_scales.scalar_type() == at::ScalarType::Float,
"scalar type of a_scales must be float");
TORCH_CHECK(global_scale.scalar_type() == c.scalar_type(),
"scalar type of global_scale must be the same with c");
TORCH_CHECK(global_scale.scalar_type() == at::ScalarType::Float,
"scalar type of global_scale must be float");
if (a_type.size_bits() == 16) {
TORCH_CHECK(
a.scalar_type() == c.scalar_type(),
+15 -10
View File
@@ -251,8 +251,8 @@ __global__ void Marlin(
const float* __restrict__ a_scales_ptr,
// fp16 quantization scales. shape (k/groupsize, n)
const int4* __restrict__ scales_ptr,
// fp16 global scale (for nvfp4// only)
const uint16_t* __restrict__ global_scale_ptr,
// float global scale (for nvfp4// only)
const float* __restrict__ global_scale_ptr,
// 4bit packed zero-points of shape
// (k/groupsize, n/pack_factor)
const int4* __restrict__ zp_ptr,
@@ -292,7 +292,13 @@ __global__ void Marlin(
#endif
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
constexpr bool use_fp16_accum = a_type_id == vllm::kFloat16.id();
constexpr auto num_bits = vllm::ScalarType::from_id(b_type_id).size_bits();
// Disable use_fp16_accum for NVFP4 and cases when group_size == -1 &&
// num_bits == 4
constexpr bool use_fp16_accum =
a_type_id == vllm::kFloat16.id() &&
(!(b_type_id == vllm::kFE2M1f.id() && s_type_id == vllm::kFE4M3fn.id()) &&
!(group_blocks == -1 && num_bits == 4));
#else
constexpr bool use_fp16_accum = false;
#endif
@@ -342,11 +348,10 @@ __global__ void Marlin(
has_zp && !is_zp_float && !std::is_same<scalar_t, nv_bfloat16>::value ||
has_zp && !is_zp_float && !(b_type == vllm::kU8);
c_scalar_t2 global_scale;
float global_scale_f32 = 1.0f;
if constexpr (b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
uint16_t val = global_scale_ptr[0];
global_scale = Cdtype::num2num2(*reinterpret_cast<c_scalar_t*>(&val));
global_scale_f32 = global_scale_ptr[0];
}
constexpr bool has_act_order = group_blocks == 0;
@@ -1644,6 +1649,10 @@ __global__ void Marlin(
// We first reorder in shared memory to guarantee the most efficient final
// global write patterns
auto write = [&](int idx, float c0, float c1, FragS& s, FragS& b_bias) {
if constexpr (b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
c0 *= global_scale_f32;
c1 *= global_scale_f32;
}
c_scalar_t2 res =
Cdtype::nums2num2(Cdtype::float2num(c0), Cdtype::float2num(c1));
@@ -1659,10 +1668,6 @@ __global__ void Marlin(
}
res = __hmul2(res, tmp_scale);
}
if constexpr (b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
res = __hmul2(res, global_scale);
}
if (has_bias && last) {
c_scalar_t2 tmp_bias = b_bias[0];
if constexpr (m_block_size_8) {