Skip to content

[https://nvbugs/6198760][fix] Refresh FMHA cubins to fix SageAttention when KV-sequence is not a multiple of 128 #17648

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
xrq-phys merged 3 commits into from
Aug 18, 2026
Merged

[https://nvbugs/6198760][fix] Refresh FMHA cubins to fix SageAttention when KV-sequence is not a multiple of 128 #17648

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c2be19b
Refresh cubins for kernel fix
xrq-phys Aug 15, 2026
bbe6a20
Unwave Sage 1560 test
xrq-phys Aug 15, 2026
724b944
Adapt SF handling for fixed SageAttn kernels: fix SeqLen%128!=0, fix …
xrq-phys Aug 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
The diff you're trying to view is too large. We only load the first 3000 changed files.
36 changes: 22 additions & 14 deletions cpp/tensorrt_llm/common/attentionOp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -881,15 +881,16 @@ size_t AttentionOp::getWorkspaceSizeForContext(tensorrt_llm::DataType type, int3
else if (useSageAttnSeparateQkv)
{
fp8_q_buf_size = max_num_tokens * static_cast<size_t>(local_hidden_units_qo);
fp8_k_buf_size = max_num_tokens * static_cast<size_t>(local_hidden_units_kv);
fp8_v_buf_size = max_num_tokens * static_cast<size_t>(local_hidden_units_kv);
fp8_k_buf_size = total_kv_len * static_cast<size_t>(local_hidden_units_kv);
fp8_v_buf_size = total_kv_len * static_cast<size_t>(local_hidden_units_kv);
}

int32_t const q_max_n_blk = mSageAttnNumEltsPerBlkQ > 0 ? tc::divUp(input_seq_length, mSageAttnNumEltsPerBlkQ) : 0;
int32_t const k_max_n_blk = mSageAttnNumEltsPerBlkK > 0 ? tc::divUp(kv_seq_length, mSageAttnNumEltsPerBlkK) : 0;
size_t const sage_q_sfs_buffer_size = sizeof(float) * mNumAttnHeads * batch_size * static_cast<size_t>(q_max_n_blk);
size_t const sage_k_sfs_buffer_size
= sizeof(float) * mNumAttnKVHeads * batch_size * static_cast<size_t>(k_max_n_blk);
int32_t const q_max_n_blk
= mSageAttnNumEltsPerBlkQ > 0 ? tc::divUp(max_num_tokens, mSageAttnNumEltsPerBlkQ) + batch_size - 1 : 0;
int32_t const k_max_n_blk
= mSageAttnNumEltsPerBlkK > 0 ? tc::divUp(total_kv_len, mSageAttnNumEltsPerBlkK) + batch_size - 1 : 0;
size_t const sage_q_sfs_buffer_size = sizeof(float) * mNumAttnHeads * static_cast<size_t>(q_max_n_blk);
size_t const sage_k_sfs_buffer_size = sizeof(float) * mNumAttnKVHeads * static_cast<size_t>(k_max_n_blk);
size_t const sage_v_sfs_buffer_size = mSageAttnNumEltsPerBlkV > 0
? sizeof(float) * tc::divUp(local_hidden_units_kv, std::max(1, mSageAttnNumEltsPerBlkV))
: 0;
Expand Down Expand Up @@ -1577,15 +1578,16 @@ int AttentionOp::enqueueContext(EnqueueContextParams<T> const& params, cudaStrea
fp8_v_buf_size = params.total_kv_len * static_cast<size_t>(local_hidden_units_kv);
}

int32_t const q_max_n_blk
= mSageAttnNumEltsPerBlkQ > 0 ? tc::divUp(params.input_seq_length, mSageAttnNumEltsPerBlkQ) : 0;
int32_t const k_max_n_blk = mSageAttnNumEltsPerBlkK > 0 ? tc::divUp(kv_seq_length, mSageAttnNumEltsPerBlkK) : 0;
int32_t const q_max_n_blk = mSageAttnNumEltsPerBlkQ > 0
? tc::divUp(params.num_tokens, mSageAttnNumEltsPerBlkQ) + params.batch_size - 1
: 0;
int32_t const k_max_n_blk = mSageAttnNumEltsPerBlkK > 0
? tc::divUp(params.total_kv_len, mSageAttnNumEltsPerBlkK) + params.batch_size - 1
: 0;
int32_t const v_max_n_blk
= mSageAttnNumEltsPerBlkV > 0 ? tc::divUp(local_hidden_units_kv, mSageAttnNumEltsPerBlkV) : 0;
size_t const sage_q_sfs_buffer_size
= sizeof(float) * mNumAttnHeads * params.batch_size * static_cast<size_t>(q_max_n_blk);
size_t const sage_k_sfs_buffer_size
= sizeof(float) * mNumAttnKVHeads * params.batch_size * static_cast<size_t>(k_max_n_blk);
size_t const sage_q_sfs_buffer_size = sizeof(float) * mNumAttnHeads * static_cast<size_t>(q_max_n_blk);
size_t const sage_k_sfs_buffer_size = sizeof(float) * mNumAttnKVHeads * static_cast<size_t>(k_max_n_blk);
size_t const sage_v_sfs_buffer_size = sizeof(float) * v_max_n_blk;

size_t const padding_offset_size
Expand Down Expand Up @@ -1905,6 +1907,8 @@ int AttentionOp::enqueueContext(EnqueueContextParams<T> const& params, cudaStrea
{
TLLM_CHECK_WITH_INFO(mFP8ContextFMHA, "SageAttention kernel runs under mFP8ContextFMHA option.");
TLLM_CHECK_WITH_INFO(mFmhaDispatcher->isSupported(), "SageAttention has no unfused fallback implemented.");
TLLM_CHECK_WITH_INFO(mMaskType == AttentionMaskType::PADDING,
"SageAttention only supports dense (padding) mask, got mask type %d.", static_cast<int>(mMaskType));
TLLM_CHECK_WITH_INFO(
mSageAttnNumEltsPerBlkQ > 0 && mSageAttnNumEltsPerBlkK > 0 && mSageAttnNumEltsPerBlkV == 1,
"SageQuant requires positive block sizes for Q and K while the block size for V must be 1.");
Expand All @@ -1928,8 +1932,10 @@ int AttentionOp::enqueueContext(EnqueueContextParams<T> const& params, cudaStrea

// Quantize into Fp8Q, SfsQ, SfsV
sageQuantParams.sumSeqLensQk = params.num_tokens;
sageQuantParams.batchSize = params.batch_size;
sageQuantParams.numHeads = mNumAttnHeads;
sageQuantParams.tokenBlockSize = mSageAttnNumEltsPerBlkQ;
sageQuantParams.ptrCuSeqLensQk = contextCuQSeqlens;
sageQuantParams.ptrQk = attention_input;
sageQuantParams.ptrQkQuant = workspaceViews.fp8QBuf;
sageQuantParams.ptrQkScale = workspaceViews.sageQScale;
Expand All @@ -1938,8 +1944,10 @@ int AttentionOp::enqueueContext(EnqueueContextParams<T> const& params, cudaStrea

// Quantize into Fp8K, SfsK, Fp8V
sageQuantParams.sumSeqLensQk = params.total_kv_len;
sageQuantParams.batchSize = params.batch_size;
sageQuantParams.numHeads = mNumAttnKVHeads;
sageQuantParams.tokenBlockSize = mSageAttnNumEltsPerBlkK;
sageQuantParams.ptrCuSeqLensQk = contextCuKvSeqlens;
Comment thread
coderabbitai[bot] marked this conversation as resolved.
sageQuantParams.ptrQk = params.k_ptr;
sageQuantParams.ptrQkQuant = workspaceViews.fp8KBuf;
sageQuantParams.ptrQkScale = workspaceViews.sageKScale;
Expand Down
Loading
Loading