Skip to content

ggml-hrx: masked KV rows reach flash-attention P*V as V = +0 (identical requests gave different logits) #72

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
bong-water-water-bong merged 1 commit into from
Oct 2, 2026
Merged

ggml-hrx: masked KV rows reach flash-attention P*V as V = +0 (identical requests gave different logits) #72

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
85 changes: 80 additions & 5 deletions ...ml-hrx/kernel-corpus/kernels/loom-libs/ops/flash_attention_decode_split_f32_f16_wmma.loom
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,7 @@ template.def<@ggml.flash_attention.decode_split.produce_partials.active> device
%c0_f32 = scalar.constant 0.0 : f32
%negative_large = scalar.constant -1e+30 : f32
%c0_i32 = scalar.constant 0 : i32
%masked_value_stage_bytes = index.constant 1024 : offset
%c1_i32 = scalar.constant 1 : i32
%attention_scale = config.get @ggml.flash_attention.attention_scale : f32
%c0_f16 = scalar.constant 0.0 : f16
Expand Down Expand Up @@ -153,6 +154,13 @@ template.def<@ggml.flash_attention.decode_split.produce_partials.active> device
%probability_stage = buffer.alloca<workgroup> align(16) %probability_stage_bytes : buffer
%product_stage = buffer.alloca<workgroup> align(16) %product_stage_bytes : buffer
%tail_value_stage = buffer.alloca<workgroup> align(16) %product_stage_bytes : buffer
%visibility_stage_bytes = index.constant 256 : offset
%visibility_stage = buffer.alloca<workgroup> align(16) %visibility_stage_bytes : buffer
%visibility_view = buffer.view %visibility_stage[%c0_offset] : buffer -> view<64xf32>
// Per-subgroup 16x32 V panels for key tiles that mix visible and masked
// keys; the score stage is free once probabilities are published.
%masked_value_stage_offset = index.scale %subgroup, %masked_value_stage_bytes : index, offset -> offset
%masked_value_stage_view = buffer.view %score_stage[%masked_value_stage_offset] : buffer -> view<16x32xf16>
%query_stage_view = buffer.view %query_stage[%c0_offset] : buffer -> view<16x[%query_stage_stride]xf16>
%query_transposed_view = buffer.view %query_stage[%c0_offset] : buffer -> view<[%qk_head_size]x16xf16, %query_transposed_layout>
%score_stage_view = buffer.view %score_stage[%c0_offset] : buffer -> view<64x24xf32>
Expand Down Expand Up @@ -274,6 +282,19 @@ template.def<@ggml.flash_attention.decode_split.produce_partials.active> device
view.store %probability1, %probability_stage_view[%query_row1, %lane] : f16, view<16x64xf16, %probability_transposed_layout>
view.store %probability2, %probability_stage_view[%query_row2, %lane] : f16, view<16x64xf16, %probability_transposed_layout>
view.store %probability3, %probability_stage_view[%query_row3, %lane] : f16, view<16x64xf16, %probability_transposed_layout>
// Masked keys must reach P*V as V = +0, not just P = 0. The F16 WMMA
// result depended on the sign of V at keys whose probability is zero
// (measured: flipping the sign of one masked V row changed one output
// element by one F16 ulp). Rows past the sequence end hold whatever an
// earlier request left there, so identical requests gave different logits.
// Key tiles with every key visible read V directly, fully masked tiles
// use a zero V fragment, and mixed tiles stage V with masked rows zeroed.
%is_first_subgroup = index.cmp eq, %subgroup, %c0 : index
scf.if %is_first_subgroup {
view.store %lane_mask, %visibility_view[%lane] : f32, view<64xf32>
}
%lane_key_visible = scalar.cmpf ogt, %lane_mask, %negative_large : f32
%lane_key_tile = index.div %lane, %c16 : index
kernel.barrier<workgroup> scope(workgroup) ordering(acq_rel)
scf.for %output_tile = [%c0 to %output_tile_count step %c1] unroll {
%output_tile_channel = index.mul %output_tile, %c128 : index
Expand Down Expand Up @@ -319,10 +340,59 @@ template.def<@ggml.flash_attention.decode_split.produce_partials.active> device
%value_token, %value_token_end = index.assume %value_token0, %value_token_end0 [le(%value_token_end0, %bounded_key_value_token_count)] : index, index
%full_value_view = buffer.view %value_aligned[%c0_offset] : buffer -> view<[%value_token_end]x[%value_width]xf16>
%probability_fragment = vector.fragment.load<lhs> %probability_stage_view[%c0, %key_tile] shape [%m, %k] : view<16x64xf16, %probability_transposed_layout> -> vector<16xf16>
%value_fragment0 = vector.fragment.load<rhs> %full_value_view[%value_token, %clamped_value_channel0] shape [%k, %n] : view<[%value_token_end]x[%value_width]xf16> -> vector<16xf16>
%value_fragment1 = vector.fragment.load<rhs> %full_value_view[%value_token, %clamped_value_channel1] shape [%k, %n] : view<[%value_token_end]x[%value_width]xf16> -> vector<16xf16>
%next_product_accumulator0 = vector.mma %probability_fragment, %value_fragment0, %product_accumulator0 : vector<16xf16>, vector<16xf16>, vector<8xf16>
%next_product_accumulator1 = vector.mma %probability_fragment, %value_fragment1, %product_accumulator1 : vector<16xf16>, vector<16xf16>, vector<8xf16>
%key_tile_ordinal = index.div %key_tile, %c16 : index
%lane_in_key_tile = index.cmp eq, %lane_key_tile, %key_tile_ordinal : index
%lane_outside_key_tile = index.cmp ne, %lane_key_tile, %key_tile_ordinal : index
%lane_tile_visible = scalar.ori %lane_outside_key_tile, %lane_key_visible : i1
%lane_tile_has_visible = scalar.andi %lane_in_key_tile, %lane_key_visible : i1
%key_tile_all_visible = kernel.subgroup.vote.all %lane_tile_visible : i1
%key_tile_any_visible = kernel.subgroup.vote.any %lane_tile_has_visible : i1
%next_product_accumulator0, %next_product_accumulator1 = scf.if %key_tile_all_visible -> (vector<8xf16>, vector<8xf16>) {
%value_fragment0 = vector.fragment.load<rhs> %full_value_view[%value_token, %clamped_value_channel0] shape [%k, %n] : view<[%value_token_end]x[%value_width]xf16> -> vector<16xf16>
%value_fragment1 = vector.fragment.load<rhs> %full_value_view[%value_token, %clamped_value_channel1] shape [%k, %n] : view<[%value_token_end]x[%value_width]xf16> -> vector<16xf16>
%direct_accumulator0 = vector.mma %probability_fragment, %value_fragment0, %product_accumulator0 : vector<16xf16>, vector<16xf16>, vector<8xf16>
%direct_accumulator1 = vector.mma %probability_fragment, %value_fragment1, %product_accumulator1 : vector<16xf16>, vector<16xf16>, vector<8xf16>
scf.yield %direct_accumulator0, %direct_accumulator1 : vector<8xf16>, vector<8xf16>
} else {
%masked_accumulator0, %masked_accumulator1 = scf.if %key_tile_any_visible -> (vector<8xf16>, vector<8xf16>) {
scf.for %stage_iteration = [%c0 to %c8 step %c1] unroll {
%stage_linear = index.madd %stage_iteration, %c64, %lane : index
%stage_key_row = index.div %stage_linear, %c32 : index
%stage_column = index.rem %stage_linear, %c32 : index
%stage_second_fragment = index.cmp uge, %stage_column, %c16 : index
%stage_fragment_column = index.rem %stage_column, %c16 : index
%stage_channel_base = scf.select %stage_second_fragment, %clamped_value_channel1, %clamped_value_channel0 : index
%stage_channel0 = index.add %stage_channel_base, %stage_fragment_column : index
%stage_channel = index.assume %stage_channel0 [lt(%stage_channel0, %value_width)] : index
%stage_block_key0 = index.add %key_tile, %stage_key_row : index
%stage_block_key = index.min %stage_block_key0, %c63 : index
%stage_key_mask = view.load %visibility_view[%stage_block_key] : view<64xf32> -> f32
%stage_key_visible = scalar.cmpf ogt, %stage_key_mask, %negative_large : f32
%stage_value = scf.if %stage_key_visible -> (f16) {
%stage_token0 = index.add %value_token, %stage_key_row : index
%stage_token = index.assume %stage_token0 [lt(%stage_token0, %value_token_end)] : index
%loaded = view.load %full_value_view[%stage_token, %stage_channel] : view<[%value_token_end]x[%value_width]xf16> -> f16
scf.yield %loaded : f16
} else {
scf.yield %c0_f16 : f16
}
view.store %stage_value, %masked_value_stage_view[%stage_key_row, %stage_column] : f16, view<16x32xf16>
}
kernel.barrier<workgroup> scope(subgroup) ordering(acq_rel)
%staged_value_fragment0 = vector.fragment.load<rhs> %masked_value_stage_view[%c0, %c0] shape [%k, %n] : view<16x32xf16> -> vector<16xf16>
%staged_value_fragment1 = vector.fragment.load<rhs> %masked_value_stage_view[%c0, %c16] shape [%k, %n] : view<16x32xf16> -> vector<16xf16>
%staged_accumulator0 = vector.mma %probability_fragment, %staged_value_fragment0, %product_accumulator0 : vector<16xf16>, vector<16xf16>, vector<8xf16>
%staged_accumulator1 = vector.mma %probability_fragment, %staged_value_fragment1, %product_accumulator1 : vector<16xf16>, vector<16xf16>, vector<8xf16>
kernel.barrier<workgroup> scope(subgroup) ordering(acq_rel)
scf.yield %staged_accumulator0, %staged_accumulator1 : vector<8xf16>, vector<8xf16>
} else {
%zero_value_fragment = vector.fragment<rhs> %c0_f16x16 shape [%k, %n] : vector<16xf16>
%zero_accumulator0 = vector.mma %probability_fragment, %zero_value_fragment, %product_accumulator0 : vector<16xf16>, vector<16xf16>, vector<8xf16>
%zero_accumulator1 = vector.mma %probability_fragment, %zero_value_fragment, %product_accumulator1 : vector<16xf16>, vector<16xf16>, vector<8xf16>
scf.yield %zero_accumulator0, %zero_accumulator1 : vector<8xf16>, vector<8xf16>
}
scf.yield %masked_accumulator0, %masked_accumulator1 : vector<8xf16>, vector<8xf16>
}
scf.yield %next_product_accumulator0, %next_product_accumulator1 : vector<8xf16>, vector<8xf16>
}
scf.yield %full_product_fragment0, %full_product_fragment1 : vector<8xf16>, vector<8xf16>
Expand All @@ -337,7 +407,12 @@ template.def<@ggml.flash_attention.decode_split.produce_partials.active> device
%value_token = index.add %value_token0, %key_row : index
%value_token_valid = index.cmp ult, %value_token, %bounded_key_value_token_count : index
%value_channel_valid = index.cmp ult, %value_channel, %value_head_size : index
%value_valid = scalar.andi %value_token_valid, %value_channel_valid : i1
%value_valid0 = scalar.andi %value_token_valid, %value_channel_valid : i1
%block_value_key0 = index.add %key_tile, %key_row : index
%block_value_key = index.min %block_value_key0, %c63 : index
%value_key_mask = view.load %visibility_view[%block_value_key] : view<64xf32> -> f32
%value_key_visible = scalar.cmpf ogt, %value_key_mask, %negative_large : f32
%value_valid = scalar.andi %value_valid0, %value_key_visible : i1
%value_element = scf.if %value_valid -> (f16) {
%global_value_channel = index.add %value_head_base, %value_channel : index
%loaded = view.load %value_view[%value_token, %global_value_channel] : view<[%bounded_key_value_token_count]x[%value_width]xf16> -> f16
Expand Down
Loading
Loading