Skip to content

Commit

Permalink
bug fixes and update Legion version
Browse files Browse the repository at this point in the history
  • Loading branch information
jiazhihao committed Jan 2, 2024
1 parent a45826e commit 093b29d
Show file tree
Hide file tree
Showing 4 changed files with 5 additions and 13 deletions.
2 changes: 1 addition & 1 deletion deps/legion
Submodule legion updated from 626b55 to d06527
8 changes: 2 additions & 6 deletions src/ops/linear.cc
Original file line number Diff line number Diff line change
Expand Up @@ -467,12 +467,8 @@ OpMeta *Linear::init_task_with_dim(Task const *task,
ctx,
runtime,
false /*readOutput*/);
TensorAccessorW<WT, NDIM> acc_kernel(regions[2],
task->regions[2],
FID_DATA,
ctx,
runtime,
false /*readOutput*/);
TensorAccessorR<WT, NDIM> acc_kernel(
regions[2], task->regions[2], FID_DATA, ctx, runtime);

// TensorAccessorR<float, 1> acc_bias(
// regions[3], task->regions[3], FID_DATA, ctx, runtime);
Expand Down
1 change: 0 additions & 1 deletion src/runtime/model.cu
Original file line number Diff line number Diff line change
Expand Up @@ -175,7 +175,6 @@ FFHandler
} else {
handle.batch_config_metadata = nullptr;
}


// checkCUDA(cudaMalloc(&handle.workSpace, handle.workSpaceSize));
#ifdef FF_USE_NCCL
Expand Down
7 changes: 2 additions & 5 deletions src/runtime/request_manager.cc
Original file line number Diff line number Diff line change
Expand Up @@ -1188,10 +1188,7 @@ BeamSearchBatchConfig
int ssm_decoding_steps =
profiling_requests[request.guid].ssm_decoding_steps;

new_bc.beamRequestsInfo[i].beam_size =
spec_infer_tree_width.size() > ssm_decoding_steps
? spec_infer_tree_width[ssm_decoding_steps]
: 1;
new_bc.beamRequestsInfo[i].beam_size = 1;
// printf("beam size: %d, %d\n",
// new_bc.beamRequestsInfo[i].beam_size,
// ssm_decoding_steps);
Expand Down Expand Up @@ -1820,7 +1817,7 @@ void RequestManager::updateBitMask(BatchConfig::BitMask &bitmask,
void RequestManager::appendPendingRequest(BatchConfig::BitMask &bitmask,
int initLength) {
assert(initLength > 0);
std::cout << "append pending bit mask: " << initLength << "\n";
// std::cout << "append pending bit mask: " << initLength << "\n";
// eg. 4 tokens: t1: 0000000..1111, t2: 0000000..1110, t3: 0000000..1100, t4:
// 0000000..1000
bitmask.non_tree_cache_size = 0;
Expand Down

0 comments on commit 093b29d

Please sign in to comment.