Skip to content

[bug fix] Issue 3700 sparse mla sm121 hang #4732

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
kahyunnam merged 3 commits into from
Sep 1, 2026
+24 −5
Merged

[bug fix] Issue 3700 sparse mla sm121 hang #4732

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
b0f7880
fix(sparse-mla): add IO-warp execution barrier in prefill KV gather
kahyunnam Aug 25, 2026
1c9a107
fix(sparse-mla): gate IO-warp barrier to sm_121
kahyunnam Aug 25, 2026
c3cacb1
fix(sparse-mla): alternate barrier ids in prefill KV handshake
kahyunnam Aug 25, 2026
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
19 changes: 19 additions & 0 deletions include/flashinfer/attention/sparse_mla_sm120/arch/barrier.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,25 @@ __device__ __forceinline__ void bar_sync_t() {
asm volatile("barrier.cta.sync %0, %1;\n" ::"n"(ID), "n"(CNT) : "memory");
}

// Producer/consumer handshakes where the arriving side does not block must alternate
// between two barrier ids: the non-blocking side may be one iteration ahead, and
// arriving twice in one phase is invalid and desynchronizes the barrier (#3700).
template <int ID_EVEN, int ID_ODD, int CNT>
__device__ __forceinline__ void bar_arrive_alt(int parity) {
if (parity)
bar_arrive_t<ID_ODD, CNT>();
else
bar_arrive_t<ID_EVEN, CNT>();
}

template <int ID_EVEN, int ID_ODD, int CNT>
__device__ __forceinline__ void bar_sync_alt(int parity) {
if (parity)
bar_sync_t<ID_ODD, CNT>();
else
bar_sync_t<ID_EVEN, CNT>();
}

// mbarrier (SM90+) for async copy tracking
__device__ __forceinline__ void mbarrier_init(uint64_t* mbar, uint32_t count) {
uint32_t addr = static_cast<uint32_t>(__cvta_generic_to_shared(mbar));
Expand Down
10 changes: 5 additions & 5 deletions include/flashinfer/attention/sparse_mla_sm120/prefill_kernel.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -154,7 +154,7 @@ __global__ void __launch_bounds__(BLOCK_THREADS, 1)
sm.kv_bufs[(ti + 1) & 1], idx_base + (ti + 1) * BI, KV_cache,
sm.mbar_kv + ((ti + 1) & 1), io_tid, stride_kv_block, kv_l2_policy);
}
bar_sync_t<1, BLOCK_THREADS>();
bar_sync_alt<1, 5, BLOCK_THREADS>(ti & 1);
}

// ── Math warps ──────────────────────────────────────────────────
Expand Down Expand Up @@ -519,7 +519,7 @@ __global__ void __launch_bounds__(BLOCK_THREADS, 1)
stride_kv_block, reinterpret_cast<bf16*>(sm.w_fp8));
}

bar_arrive_t<1, BLOCK_THREADS>();
bar_arrive_alt<1, 5, BLOCK_THREADS>(ti & 1);
if (ti + 1 < actual_ni) {
const int next_phase = ((ti + 1) >> 1) & 1;
mbarrier_wait_parity(sm.mbar_kv + ((ti + 1) & 1), next_phase);
Expand Down Expand Up @@ -782,7 +782,7 @@ __device__ __forceinline__ void prefill_mg_impl(
io_tid, stride_kv_block, kv_l2_policy);
}
}
bar_sync_t<1, BLOCK_THREADS>();
bar_sync_alt<1, 5, BLOCK_THREADS>(ti & 1);
}
} else {
auto issue_tile = [&](int logical_ti, int buf) {
Expand Down Expand Up @@ -823,7 +823,7 @@ __device__ __forceinline__ void prefill_mg_impl(
if (ti + 1 < loop_bound) {
issue_tile(ti + 1, (ti + 1) & 1);
}
bar_sync_t<1, BLOCK_THREADS>();
bar_sync_alt<1, 5, BLOCK_THREADS>(ti & 1);
}
}

Expand Down Expand Up @@ -1514,7 +1514,7 @@ __device__ __forceinline__ void prefill_mg_impl(
reinterpret_cast<bf16*>(sm.w_fp8()));
}
}
bar_arrive_t<1, BLOCK_THREADS>();
bar_arrive_alt<1, 5, BLOCK_THREADS>(ti & 1);
if (ti + 1 < loop_bound) {
const int next_phase = ((ti + 1) >> 1) & 1;
mbarrier_wait_parity(sm.mbar_kv((ti + 1) & 1), next_phase);
Expand Down
Loading