Skip to content

Magpie/1.1.0 #9

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
anand-nv merged 3 commits into from
Nov 13, 2025
Merged

Magpie/1.1.0 #9

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
0142d4a
Fix missing attention prior kernels
anand-nv Nov 12, 2025
0535805
Code cleanup
anand-nv Nov 12, 2025
05cc1e2
Fix missing check for getting stuck on last idx
anand-nv Nov 13, 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
4 changes: 2 additions & 2 deletions cpp/tensorrt_llm/batch_manager/trtGptModelInflightBatching.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2107,7 +2107,6 @@ runtime::CudaEvent TrtGptModelInflightBatching::decoderStepAsync(ScheduledReques
for (SizeType32 vid = 0; vid < getNumVocabs(); vid++)
{
auto& decoderInputBuffers = mDecoderInputBuffers.at(getFusedBufferId());
auto& decoderOutputBuffers = mDecoderOutputBuffers[vid].at(getFusedBufferId());
auto& decoderState = mDecoderStates[vid];

auto const contextBufferId = mCtxGenFusion ? getFusedBufferId() : getContextBufferId();
Expand Down Expand Up @@ -2433,7 +2432,8 @@ void TrtGptModelInflightBatching::updateRequests(ScheduledRequests const& schedu

// Terminate if request has finished or if it is speculative decoding target model
if (decoderFinishedSumPtr[seqSlot] == reqBeamWidth
|| (mModelConfig.getSpeculativeDecodingMode().isDraftTokensExternal() && llmReq->hasDraftTokens()))
|| (mModelConfig.getSpeculativeDecodingMode().isDraftTokensExternal() && llmReq->hasDraftTokens())
|| (mModelConfig.useAttentionPrior() && llmReq->isAttentionPriorFinished()))
{
postProcessRequest(*llmReq, numDroppedTokens);

Expand Down
56 changes: 53 additions & 3 deletions ...rrt_llm/kernels/decoderMaskedMultiheadAttention/decoderMaskedMultiheadAttentionTemplate.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1528,6 +1528,18 @@ __global__ void __launch_bounds__(MAX_THEADS_PER_BLOCK, MIN_BLOCKS_PER_SM) maske
float const kv_scale_quant_orig_f = (ENABLE_8BITS_KV_CACHE ? params.kv_scale_quant_orig[0] : 1.0f);
convert_from_float(&k_scale_quant_orig, k_scale_quant_orig_f);
convert_from_float(&kv_scale_orig_quant, (ENABLE_8BITS_KV_CACHE ? params.kv_scale_orig_quant[0] : 1.0f));
// parameters related to attention prior
int focus;
if (params.attention_prior_focus != nullptr)
{
focus = params.attention_prior_focus[batch_beam_idx];
}
bool const store_scores = params.attention_prior_scores != nullptr;
float* scores_ptr = nullptr;
if (store_scores)
{
scores_ptr = &params.attention_prior_scores[batch_beam_idx * params.attention_prior_lookahead];
}

#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaGridDependencySynchronize();
Expand Down Expand Up @@ -1849,7 +1861,7 @@ __global__ void __launch_bounds__(MAX_THEADS_PER_BLOCK, MIN_BLOCKS_PER_SM) maske
{
relative_attention_bias = convert_to_float(relative_attention_bias_ptr[tlength]);
}
if (has_attention_mask && tidx == 0)
if (has_attention_mask && tidx == 0 && !DO_CROSS_ATTENTION)
{
// Note: reuse the relative_attention_bias variable.
// attention_mask = 1.0 means that the position is not masked.
Expand Down Expand Up @@ -2055,7 +2067,7 @@ __global__ void __launch_bounds__(MAX_THEADS_PER_BLOCK, MIN_BLOCKS_PER_SM) maske
{
relative_attention_bias = convert_to_float(relative_attention_bias_ptr[local_time_now]);
}
if (is_active && has_attention_mask)
if (is_active && has_attention_mask && !DO_CROSS_ATTENTION)
{
// Note: reuse the relative_attention_bias variable.
// attention_mask = 1.0 means that the position is not masked.
Expand Down Expand Up @@ -2268,13 +2280,30 @@ __global__ void __launch_bounds__(MAX_THEADS_PER_BLOCK, MIN_BLOCKS_PER_SM) maske
float inv_sum = __fdividef(logit_scale, sum + 1.e-6f);

int const normlization_loop_end = MULTI_BLOCK_FLAG ? timesteps_per_block : kv_loop_length;
float sum_rescale = 0.0f;
for (int ti = tidx; ti <= normlization_loop_end; ti += THREADS_PER_BLOCK)
{
int const time_now = MULTI_BLOCK_FLAG ? ti + c_tile_times_timesteps_per_block : ti;

if (!MULTI_BLOCK_FLAG)
{
convert_from_float(&logits_smem[ti], qk_smem[ti] * inv_sum);
float prob = qk_smem[ti] * inv_sum;
if (DO_CROSS_ATTENTION && params.attention_prior_focus != nullptr)
{
// do the masking to the prob
if (ti < (focus - params.attention_prior_window_left)
|| ti > (focus + params.attention_prior_window_right))
{
prob *= 0.1f;
}
// store back
qk_smem[ti] = prob;
sum_rescale += prob;
}
else
{
convert_from_float(&logits_smem[ti], prob);
}
}
else
{
Expand All @@ -2290,6 +2319,27 @@ __global__ void __launch_bounds__(MAX_THEADS_PER_BLOCK, MIN_BLOCKS_PER_SM) maske
}
}

// for the case when we apply prior, we need to perform additional normalization,
// dividing by the sum of the modified probs.
__syncthreads();
if (!MULTI_BLOCK_FLAG && DO_CROSS_ATTENTION && params.attention_prior_focus != nullptr)
{
sum_rescale = block_sum<WARPS_PER_BLOCK>(&red_smem[WARPS_PER_BLOCK], sum_rescale);

// finally loop to compute probability, store probability to buffer if needed
float inv_sum_rescale = __fdividef(1.0f, sum_rescale + 1.e-6f);
for (int ti = tidx; ti <= kv_loop_length; ti += THREADS_PER_BLOCK)
{
float prob = qk_smem[ti] * inv_sum_rescale;
if (store_scores && ti >= focus && ti < focus + params.attention_prior_lookahead)
{
scores_ptr[ti - focus] = prob;
}
convert_from_float(&logits_smem[ti], prob);
}
__syncthreads();
}

// Put Values part below so we leverage __syncthreads
// from the previous step

Expand Down
Loading