Skip to content

[BugFix] Accuracy fix for llama4 int4 - improperly casted scales #16801

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Apr 18, 2025
Merged

[BugFix] Accuracy fix for llama4 int4 - improperly casted scales #16801

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1263b42
fixes for llama4 int4
LucasWilkinson Apr 17, 2025
ca7f666
use float scales
LucasWilkinson Apr 17, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 3 additions & 7 deletions csrc/moe/moe_wna16.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,6 @@
template <typename scalar_t, int bit, int GROUPS>
__global__ void moe_wna16_gemm_kernel(
const scalar_t* __restrict__ input, scalar_t* __restrict__ output,

const uint32_t* __restrict__ qweight, const scalar_t* __restrict__ scales,
const uint32_t* __restrict__ qzeros,

Expand Down Expand Up @@ -54,8 +53,6 @@ __global__ void moe_wna16_gemm_kernel(
if (token_index / top_k >= size_m) break;

num_valid_tokens = m + 1;
if (blockIdx.z == 0 && offset_n < size_n)
output[token_index * size_n + offset_n] = Dtype::int2num(0);

if (expert_id != -1) {
int k_per_thread = DIVIDE(BLOCK_SIZE_K, BLOCK_SIZE_N);
Expand Down Expand Up @@ -284,8 +281,7 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
int64_t BLOCK_SIZE_K, int64_t bit) {
const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
auto options =
torch::TensorOptions().dtype(input.dtype()).device(input.device());
output.zero_();

const int num_experts = b_qweight.size(0);
const int size_m = input.size(0);
Expand All @@ -302,9 +298,9 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
const uint32_t* b_qzeros_ptr;
if (b_qzeros.has_value())
b_qzeros_ptr = (const uint32_t*)b_qzeros.value().data_ptr<uint8_t>();
const float* topk_weights_ptr;
const float* topk_weights_ptr = nullptr;
if (topk_weights.has_value())
topk_weights_ptr = (const float*)topk_weights.value().data_ptr();
topk_weights_ptr = (const float*)topk_weights.value().data_ptr<float>();

int groups_per_block_row = BLOCK_SIZE_K / group_size;
TORCH_CHECK(bit == 4 || bit == 8, "bit must be 4 or 8");
Expand Down
1 change: 1 addition & 0 deletions vllm/model_executor/layers/fused_moe/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -422,6 +422,7 @@ def __init__(

if params_dtype is None:
params_dtype = torch.get_default_dtype()
self.params_dtype = params_dtype

# Note: here we guard against accessing the TP and DP groups when
# uninitialized (this happens when testing)
Expand Down
4 changes: 2 additions & 2 deletions vllm/model_executor/models/llama4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,8 +51,8 @@ def custom_routing_function(
renormalize: bool,
) -> Tuple[torch.Tensor, torch.Tensor]:
router_scores, router_indices = fast_topk(gating_output, topk, dim=-1)
router_scores = torch.sigmoid(router_scores.float()).to(
hidden_states.dtype)
# psuedo-standard is that the router scores are floats
router_scores = torch.sigmoid(router_scores.float())
return (router_scores, router_indices.to(torch.int32))

def __init__(self,
Expand Down