a2 Device Constraints【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skillsRead this file when writing a kernel targeting a2 (easyasc.a2, deviceb3). Do not read it for a5 kernels — the two architectures differ significantly.GoalCapture all a2-specific differences from a5 so that:a5 patterns are not blindly reused on a2the correct data path, buffer, and vec model is chosen from the start1. Hardware budgets and missing featuresFor exact per-device capacities (L0A,L0B,L0C,UB,L1, cube core count, vec sub-blocks per core) seeagent/references/facts-device-runtime.md. For the a2-missing a5 features (vf,Reg/RegList/MaskReg,l0c_to_ub,ub_to_l1_nd2nz,ub_to_l1_nz,micro,l0c_to_l1(float)), seeagent/references/facts-authoring.md.Key consequence for tile strategy: a5 tile strategies that fit L0C at 256 KB overflow on a2. Always verifyTILE_M * TILE_N * 4 * 2 128 KBfor float L0C DBuff.2. Cube → vec data path on a2Sincel0c_to_ubis absent andl0c_to_l1(float)is blocked:Mandatory path:L0C → GM workspace → UBCube:l0c_to_gm_nz2ndwrites float L0C to a GM workspace buffer (FIX pipe)Vec:gm_to_ub_padreads from GM workspace into UB (MTE2 pipe)CvMutex configuration:CvMutex(0, src_end_pipePipe.FIX, dst_end_pipePipe.MTE2)This differs from the a5 standard (dst_end_pipePipe.V) because the vec sides first consumer operation isgm_to_ub_padon MTE2, not a V-pipe compute.GM workspace design:Usesplit_workspace(DT.float, [GetCubeNum(), 2, TILE_M, TILE_N])for pingpongDimension2provides double-buffering slotsIndex asws[GetCubeIdx(), slot, row_slice, col_slice]Cube writes fullTILE_Mrows; each vec sub-block readsTILE_M // 2rows3. Vec → cube data path on a2Sinceub_to_l1_nd2nzandub_to_l1_nzare a5-only, a2 cannot publish vec output directly fromUBtoL1.Mandatory path for delayed vec - cube reuse:UB - GM workspace - L1Vec:ub_to_gm_padwrites the vec result into a GM workspace buffer (MTE3 pipe)Cube:gm_to_l1_nd2nzreloads that workspace tile intoL1(MTE2 pipe)Cube then continues with the normall1_to_l0 - mmadpathThis is the stable bridge for patterns such as:stage 1 cube computes scorevec computesp_jstage 2 cube consumes delayedp_j v_jRecommended synchronization:VcMutex(1, src_end_pipePipe.MTE3, dst_end_pipePipe.FIX)Whydst_end_pipePipe.FIX:vec producer truly ends onub_to_gm_pad(MTE3)for delayed-consumer kernels, conservative release is simpler and saferthe cube side may reload from GM onMTE2, then continue throughL1 - L0 - MMAD - FIXfreeing only after the cube stage finishes avoids premature workspace reuseWorkspace design mirrors the cube - vec bridge:usesplit_workspace(dtype, [GetCubeNum(), 2, TILE_M, TILE_N])vec sub-block 0 writes rows[0:HALF_M]vec sub-block 1 writes rows[HALF_M:TILE_M]cube waits on theVcMutex, then reloads the full tile from the same slotImportant synchronization fact from the simulator/runtime model:cube-sidewait_vec()succeeds only afterbothvec lanes have produced their tokensthis makes a full-tile cube reload safe after the two half-row vec writes complete3a. Cube-side matmul dependency reuse rule on a2The same high-level rule as a5 still applies on a2:if one cube matmul only feeds a later cube matmul, prefer keeping that dependency on the cube-side path instead of inventing a UB detourOn a2 the practical boundary is stricter:l0c_to_l1existsbutl0c_to_l1(float)is blocked onb*devicesandl0c_to_ubis absentSo the stable rule is:if the intermediate dependency can be republished toL1in a supported dtype, preferL0C - L1 - L0 - mmaddonotroute a pure cube-side dependency throughUBonly fall back to GM workspace when the dependency truly needs a vec-side stage or when the requiredL1destination dtype is unsupportedPractical implication:a2 still benefits from the same avoid unnecessaryL0C - UB - L1thinking rulethe difference from a5 is not the existence ofl0c_to_l1, but the narrower dtype surface4. Sub-block execution modelEach cube core has 2 vec sub-blocks. On a2:Each sub-block has its ownindependent 192 KB UBVec instructions in the kernel body execute on both sub-blocks simultaneouslyUseGetSubBlockIdx()to compute different GM offsets for each sub-blockEach sub-block reads its ownTILE_M // 2rows from the workspaceTypical pattern:sb GetSubBlockIdx() sb_row Var(sb * HALF_M) ub_data workspace[cube_idx, slot, sb_row:sb_row HALF_M, 0:TILE_N] # ... vec processing ... out_row Var(q_row sb_row) output[out_row:out_row HALF_M, col:col TILE_N] ub_out5. Vec operations available on a2All vec operations work on UB tensors directly (no Reg intermediate):CategoryOperationsUnaryexp,ln,abs,rec,sqrt,rsqrt,reluBinaryadd,sub,mul,div,vmax,vminScalaradds,muls,vmaxs,vmins,axpyReductioncmax,cgmax,cmin,cgmin,cadd,cgadd,cpaddBroadcastdup,brcbDatamovegm_to_ub_pad,ub_to_gm_pad,ub_to_ubCastcastComparecompare,compare_scalar,set_cmpmaskSelectselectMaskset_mask,reset_mask5.1 Bitpackeduint8masks and scalar-fillselectFor a2 vec-side masking that is naturally bitpacked by column prefix, a practical shape isTensor(DT.uint8, [HALF_M, TILE_N // 8], Position.UB).Stable rules:compare(...)/compare_scalar(...)produce packed-bituint8masks for this path, andselect(...)consumes the same packed formdonottreatselectcontrol as an expanded[HALF_M, TILE_N]byte-per-element0/1tensor when the hardware contract is bitpackedfor rowwise prefix or suffix masks, prefer synthesizing mask bytes withcompare_scalar(...)against a reusable integer column-index tensor instead of populating every mask byte withSetValueTo(...)dupdoesnotsupportuint8, but whole-mask initialization can often be done efficiently through a width-compatible integer reinterpret such asmask.reinterpret(DT.int)SetValueTostill writes only the first element of the destination view, so per-byte loops are correct as a fallback but expensive in simulator control flowselect(..., SelectMode.TENSOR_SCALAR)only uses the first scalar value fromsrc2for score-domain invalidation, a small scalar-fill source such asTensor(DT.float, [1, 64], Position.UB)initialized withdup(..., neg_large)is sufficient;src2does not need to match the full score tile shapeUse this when:the control comes from an explicit packeduint8mask tensorthe fallback branch is a single scalar fill value likeneg_largeor0.0you want to avoid allocating a full-size fallback tensor just forselectyou want to amortize mask construction by building it once and reusing it across many vec operationsDiagonal causal reuse rule:for repeated diagonal masking on[HALF_M, TILE_N]tiles, it is valid to build one packed maskTensor(DT.uint8, [HALF_M, TILE_N // 8], Position.UB)and apply one full-tileselect(...)if the diagonal geometry is stable across iterations, prebuild that packed mask once and reuse itfor a2 dense backward causal tails, full128x128q-tiles have stable subblock geometry and can reuse one static packed diagonal mask perGetSubBlockIdx()tailMtiles cannot blindly reuse that full-tile mask when the kernel splits rows withhalf_rows CeilDiv(valid_m, 2), because the second subblockrow_beginshifts away from the fixed64in that case, keep the same packed-mask one-select(...)pattern, but rebuild the packed mask dynamically only for the tail tileColumn-index template trick:when the packed mask compares against contiguous column ids, build a reusable integer template once, e.g.Tensor(DT.int, [1, TILE_N], Position.UB)a compact a2 pattern is to reinterpret that tensor asint64and write twoint32column ids per store while initializing[0, 1, 2, ..., TILE_N - 1]latercompare_scalar(mask[row:row 1, 0:TILE_N // 8], col_idx, valid_cols, CompareMode.LT)synthesizes the whole row mask without scalar-per-byte loops6. UB initialization withdupOn a2, UB contents are undefined at kernel entry. There is no zero-initialization guarantee. Operations likemuls(ub, ub, 0.0)are unreliable on uninitialized buffers because0.0 × NaN NaN.Usedup(tensor, scalar_value)to fill a UB tensor with a known value.For the current validated a2 softmax kernels, a sufficiently negative finite sentinel such asneg_large -1.0e30is the default way to materialize score-domain invalidation and running-max identity state. Treat literalfloat(-inf)as simulator-convenient but hardware-fragile.dupsignature:dup(dst: Tensor, value: Union[int, float, Var], ...)Thedupoperation uses the same stride inference as other vec operations. It fills blocks and repeats according toinfer_repeat(dst)and the auto-inferred strides.Coverage analysis for common shapes (float, C08):Shaperepeatblk_strideElements coveredBuffer sizeComplete?[64, 1]111×8×8 6464✓[64, 8]808×1×8 64512✗ (only 1/8)[64, 64]64164×8×8 40964096✓[64, 128]1281128×8×8 81928192✓Warning:dupon[M, 8](broadcast-format) only fills 64 out of 512 elements. This is acceptable if the tensor is only consumed viablk_stride0operations (likesub), which also only read those 64 positions. But it is incorrect if you later attempt a full element-wise operation over the entire 512-element buffer.Practical rule: initialize in the natural computation format ([M, 1]for scalars,[M, N]for data), not in the broadcast format.Placement withinauto_syncAdupplaced inside the outer loop but before the inner loop (e.g. for per-M-tile reinitialization) is safe. It generates an extra V→MTE3 autosync event pair because auto_sync sees the V-pipedupand the inner loops MTE3ub_to_gm_padas a producer-consumer pair within the same scope.This extra event is harmless:dupcompletes before the inner loop starts, so the MTE3 event is already satisfied when the first store executes. The inner loops own V→MTE3 events for the actualexp → cast → storeflow are managed separately with different sync-key groups.with auto_sync(): for gmt in range(mt_begin, mt_end): # ... variable declarations ... dup(ub_rmax_s, neg_large) # V-pipe, safe here for nt in range(0, tiles_n): # ... cube vec tile processing ...7. Scalar computationscalar_sqrt(Var)requiresVar.dtype Datatype.float; integer Vars will TypeErrorFor1/sqrt(D), prefer passing the precomputed float value as a kernel parametermuls(ub, ub, scale_var)accepts Var(float) as the scalar argument8. Copying scalar-format UB stateFor row-scalar buffers on a2, the natural vec format is[M, 1]. This format is safe for vec binary ops likevmax,add,sub, andexp, but it isnota safe format forub_to_ubsnapshots.Why:ub_to_ubinfers burst length in units ofC0blocksfor float[64, 1], that implies copying one full 8-element block per rowthis silently corrupts or repeats scalar row state instead of copying one scalar per rowPractical rule:if you need to snapshot[M,1]running state such asprev_row_max, copy it with a vec binary op and an explicit zero buffer:dup(ub_zero_s, 0.0) add(tmp_scalar_buf, ub_rmax_s, ub_zero_s)then update or transformtmp_scalar_bufwith more vec opsdo not useub_to_ubas a generic copy for[M,1]scalar stateFiles to studyagent/example/kernels/a2/qk_matmul_batched.py— cube-only a2 baselineagent/example/kernels/a2/flash_attn_score.py— cube → vec with GM workspace bridgeagent/example/kernels/a2/flash_attn_score_iter.py— running max withdupinitialization and[M,1]vmaxagent/example/kernels/a2/flash_attn_unnorm.py— delayedexpdiff final numerator accumulation on a2agent/example/kernels/a2/flash_attn_full.py— delayed numerator accumulation plus running sum/final divide on a2agent/example/kernels/a2/flash_attn_full_pj_hif8_causal.py— rowwise diagonal-tile causal masking plusactive_tiles_nskip on a2, now with shared vec-sideDBuffscratch for stage-1/2 overlapagent/example/kernels/a2/flash_attn_full_pj_half_block32_causal.py— block-32 causal masking plusactive_tiles_nskip on a2, also using the shared vec-sideDBuffscratch lineageeasyasc/stub_functions/vec/dupbrcb.py— dup stub (validates dst is UB, infers repeat)easyasc/simulator_v2/ops/vec/v.pyandeasyasc/simulator_v2/ops/vec/_legacy_vpipe.py— current vec runtime path for dup-style execution【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
CANN/cannbot-skills a2设备约束
发布时间：2026/7/12 2:51:15
a2 Device Constraints【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skillsRead this file when writing a kernel targeting a2 (easyasc.a2, deviceb3). Do not read it for a5 kernels — the two architectures differ significantly.GoalCapture all a2-specific differences from a5 so that:a5 patterns are not blindly reused on a2the correct data path, buffer, and vec model is chosen from the start1. Hardware budgets and missing featuresFor exact per-device capacities (L0A,L0B,L0C,UB,L1, cube core count, vec sub-blocks per core) seeagent/references/facts-device-runtime.md. For the a2-missing a5 features (vf,Reg/RegList/MaskReg,l0c_to_ub,ub_to_l1_nd2nz,ub_to_l1_nz,micro,l0c_to_l1(float)), seeagent/references/facts-authoring.md.Key consequence for tile strategy: a5 tile strategies that fit L0C at 256 KB overflow on a2. Always verifyTILE_M * TILE_N * 4 * 2 128 KBfor float L0C DBuff.2. Cube → vec data path on a2Sincel0c_to_ubis absent andl0c_to_l1(float)is blocked:Mandatory path:L0C → GM workspace → UBCube:l0c_to_gm_nz2ndwrites float L0C to a GM workspace buffer (FIX pipe)Vec:gm_to_ub_padreads from GM workspace into UB (MTE2 pipe)CvMutex configuration:CvMutex(0, src_end_pipePipe.FIX, dst_end_pipePipe.MTE2)This differs from the a5 standard (dst_end_pipePipe.V) because the vec sides first consumer operation isgm_to_ub_padon MTE2, not a V-pipe compute.GM workspace design:Usesplit_workspace(DT.float, [GetCubeNum(), 2, TILE_M, TILE_N])for pingpongDimension2provides double-buffering slotsIndex asws[GetCubeIdx(), slot, row_slice, col_slice]Cube writes fullTILE_Mrows; each vec sub-block readsTILE_M // 2rows3. Vec → cube data path on a2Sinceub_to_l1_nd2nzandub_to_l1_nzare a5-only, a2 cannot publish vec output directly fromUBtoL1.Mandatory path for delayed vec - cube reuse:UB - GM workspace - L1Vec:ub_to_gm_padwrites the vec result into a GM workspace buffer (MTE3 pipe)Cube:gm_to_l1_nd2nzreloads that workspace tile intoL1(MTE2 pipe)Cube then continues with the normall1_to_l0 - mmadpathThis is the stable bridge for patterns such as:stage 1 cube computes scorevec computesp_jstage 2 cube consumes delayedp_j v_jRecommended synchronization:VcMutex(1, src_end_pipePipe.MTE3, dst_end_pipePipe.FIX)Whydst_end_pipePipe.FIX:vec producer truly ends onub_to_gm_pad(MTE3)for delayed-consumer kernels, conservative release is simpler and saferthe cube side may reload from GM onMTE2, then continue throughL1 - L0 - MMAD - FIXfreeing only after the cube stage finishes avoids premature workspace reuseWorkspace design mirrors the cube - vec bridge:usesplit_workspace(dtype, [GetCubeNum(), 2, TILE_M, TILE_N])vec sub-block 0 writes rows[0:HALF_M]vec sub-block 1 writes rows[HALF_M:TILE_M]cube waits on theVcMutex, then reloads the full tile from the same slotImportant synchronization fact from the simulator/runtime model:cube-sidewait_vec()succeeds only afterbothvec lanes have produced their tokensthis makes a full-tile cube reload safe after the two half-row vec writes complete3a. Cube-side matmul dependency reuse rule on a2The same high-level rule as a5 still applies on a2:if one cube matmul only feeds a later cube matmul, prefer keeping that dependency on the cube-side path instead of inventing a UB detourOn a2 the practical boundary is stricter:l0c_to_l1existsbutl0c_to_l1(float)is blocked onb*devicesandl0c_to_ubis absentSo the stable rule is:if the intermediate dependency can be republished toL1in a supported dtype, preferL0C - L1 - L0 - mmaddonotroute a pure cube-side dependency throughUBonly fall back to GM workspace when the dependency truly needs a vec-side stage or when the requiredL1destination dtype is unsupportedPractical implication:a2 still benefits from the same avoid unnecessaryL0C - UB - L1thinking rulethe difference from a5 is not the existence ofl0c_to_l1, but the narrower dtype surface4. Sub-block execution modelEach cube core has 2 vec sub-blocks. On a2:Each sub-block has its ownindependent 192 KB UBVec instructions in the kernel body execute on both sub-blocks simultaneouslyUseGetSubBlockIdx()to compute different GM offsets for each sub-blockEach sub-block reads its ownTILE_M // 2rows from the workspaceTypical pattern:sb GetSubBlockIdx() sb_row Var(sb * HALF_M) ub_data workspace[cube_idx, slot, sb_row:sb_row HALF_M, 0:TILE_N] # ... vec processing ... out_row Var(q_row sb_row) output[out_row:out_row HALF_M, col:col TILE_N] ub_out5. Vec operations available on a2All vec operations work on UB tensors directly (no Reg intermediate):CategoryOperationsUnaryexp,ln,abs,rec,sqrt,rsqrt,reluBinaryadd,sub,mul,div,vmax,vminScalaradds,muls,vmaxs,vmins,axpyReductioncmax,cgmax,cmin,cgmin,cadd,cgadd,cpaddBroadcastdup,brcbDatamovegm_to_ub_pad,ub_to_gm_pad,ub_to_ubCastcastComparecompare,compare_scalar,set_cmpmaskSelectselectMaskset_mask,reset_mask5.1 Bitpackeduint8masks and scalar-fillselectFor a2 vec-side masking that is naturally bitpacked by column prefix, a practical shape isTensor(DT.uint8, [HALF_M, TILE_N // 8], Position.UB).Stable rules:compare(...)/compare_scalar(...)produce packed-bituint8masks for this path, andselect(...)consumes the same packed formdonottreatselectcontrol as an expanded[HALF_M, TILE_N]byte-per-element0/1tensor when the hardware contract is bitpackedfor rowwise prefix or suffix masks, prefer synthesizing mask bytes withcompare_scalar(...)against a reusable integer column-index tensor instead of populating every mask byte withSetValueTo(...)dupdoesnotsupportuint8, but whole-mask initialization can often be done efficiently through a width-compatible integer reinterpret such asmask.reinterpret(DT.int)SetValueTostill writes only the first element of the destination view, so per-byte loops are correct as a fallback but expensive in simulator control flowselect(..., SelectMode.TENSOR_SCALAR)only uses the first scalar value fromsrc2for score-domain invalidation, a small scalar-fill source such asTensor(DT.float, [1, 64], Position.UB)initialized withdup(..., neg_large)is sufficient;src2does not need to match the full score tile shapeUse this when:the control comes from an explicit packeduint8mask tensorthe fallback branch is a single scalar fill value likeneg_largeor0.0you want to avoid allocating a full-size fallback tensor just forselectyou want to amortize mask construction by building it once and reusing it across many vec operationsDiagonal causal reuse rule:for repeated diagonal masking on[HALF_M, TILE_N]tiles, it is valid to build one packed maskTensor(DT.uint8, [HALF_M, TILE_N // 8], Position.UB)and apply one full-tileselect(...)if the diagonal geometry is stable across iterations, prebuild that packed mask once and reuse itfor a2 dense backward causal tails, full128x128q-tiles have stable subblock geometry and can reuse one static packed diagonal mask perGetSubBlockIdx()tailMtiles cannot blindly reuse that full-tile mask when the kernel splits rows withhalf_rows CeilDiv(valid_m, 2), because the second subblockrow_beginshifts away from the fixed64in that case, keep the same packed-mask one-select(...)pattern, but rebuild the packed mask dynamically only for the tail tileColumn-index template trick:when the packed mask compares against contiguous column ids, build a reusable integer template once, e.g.Tensor(DT.int, [1, TILE_N], Position.UB)a compact a2 pattern is to reinterpret that tensor asint64and write twoint32column ids per store while initializing[0, 1, 2, ..., TILE_N - 1]latercompare_scalar(mask[row:row 1, 0:TILE_N // 8], col_idx, valid_cols, CompareMode.LT)synthesizes the whole row mask without scalar-per-byte loops6. UB initialization withdupOn a2, UB contents are undefined at kernel entry. There is no zero-initialization guarantee. Operations likemuls(ub, ub, 0.0)are unreliable on uninitialized buffers because0.0 × NaN NaN.Usedup(tensor, scalar_value)to fill a UB tensor with a known value.For the current validated a2 softmax kernels, a sufficiently negative finite sentinel such asneg_large -1.0e30is the default way to materialize score-domain invalidation and running-max identity state. Treat literalfloat(-inf)as simulator-convenient but hardware-fragile.dupsignature:dup(dst: Tensor, value: Union[int, float, Var], ...)Thedupoperation uses the same stride inference as other vec operations. It fills blocks and repeats according toinfer_repeat(dst)and the auto-inferred strides.Coverage analysis for common shapes (float, C08):Shaperepeatblk_strideElements coveredBuffer sizeComplete?[64, 1]111×8×8 6464✓[64, 8]808×1×8 64512✗ (only 1/8)[64, 64]64164×8×8 40964096✓[64, 128]1281128×8×8 81928192✓Warning:dupon[M, 8](broadcast-format) only fills 64 out of 512 elements. This is acceptable if the tensor is only consumed viablk_stride0operations (likesub), which also only read those 64 positions. But it is incorrect if you later attempt a full element-wise operation over the entire 512-element buffer.Practical rule: initialize in the natural computation format ([M, 1]for scalars,[M, N]for data), not in the broadcast format.Placement withinauto_syncAdupplaced inside the outer loop but before the inner loop (e.g. for per-M-tile reinitialization) is safe. It generates an extra V→MTE3 autosync event pair because auto_sync sees the V-pipedupand the inner loops MTE3ub_to_gm_padas a producer-consumer pair within the same scope.This extra event is harmless:dupcompletes before the inner loop starts, so the MTE3 event is already satisfied when the first store executes. The inner loops own V→MTE3 events for the actualexp → cast → storeflow are managed separately with different sync-key groups.with auto_sync(): for gmt in range(mt_begin, mt_end): # ... variable declarations ... dup(ub_rmax_s, neg_large) # V-pipe, safe here for nt in range(0, tiles_n): # ... cube vec tile processing ...7. Scalar computationscalar_sqrt(Var)requiresVar.dtype Datatype.float; integer Vars will TypeErrorFor1/sqrt(D), prefer passing the precomputed float value as a kernel parametermuls(ub, ub, scale_var)accepts Var(float) as the scalar argument8. Copying scalar-format UB stateFor row-scalar buffers on a2, the natural vec format is[M, 1]. This format is safe for vec binary ops likevmax,add,sub, andexp, but it isnota safe format forub_to_ubsnapshots.Why:ub_to_ubinfers burst length in units ofC0blocksfor float[64, 1], that implies copying one full 8-element block per rowthis silently corrupts or repeats scalar row state instead of copying one scalar per rowPractical rule:if you need to snapshot[M,1]running state such asprev_row_max, copy it with a vec binary op and an explicit zero buffer:dup(ub_zero_s, 0.0) add(tmp_scalar_buf, ub_rmax_s, ub_zero_s)then update or transformtmp_scalar_bufwith more vec opsdo not useub_to_ubas a generic copy for[M,1]scalar stateFiles to studyagent/example/kernels/a2/qk_matmul_batched.py— cube-only a2 baselineagent/example/kernels/a2/flash_attn_score.py— cube → vec with GM workspace bridgeagent/example/kernels/a2/flash_attn_score_iter.py— running max withdupinitialization and[M,1]vmaxagent/example/kernels/a2/flash_attn_unnorm.py— delayedexpdiff final numerator accumulation on a2agent/example/kernels/a2/flash_attn_full.py— delayed numerator accumulation plus running sum/final divide on a2agent/example/kernels/a2/flash_attn_full_pj_hif8_causal.py— rowwise diagonal-tile causal masking plusactive_tiles_nskip on a2, now with shared vec-sideDBuffscratch for stage-1/2 overlapagent/example/kernels/a2/flash_attn_full_pj_half_block32_causal.py— block-32 causal masking plusactive_tiles_nskip on a2, also using the shared vec-sideDBuffscratch lineageeasyasc/stub_functions/vec/dupbrcb.py— dup stub (validates dst is UB, infers repeat)easyasc/simulator_v2/ops/vec/v.pyandeasyasc/simulator_v2/ops/vec/_legacy_vpipe.py— current vec runtime path for dup-style execution【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
