submission 661294

  #!POPCORN leaderboard amd-mixed-mla
  #!POPCORN gpu MI355X
- """V1261: Best of v012g safety + v1260 ps=8 speed.
- NP only for 4×1K, 32×1K (safest, small batch, safe reduce patched).
- Persistent wrapper for 64×1K, 256×1K (from v012g — avoids NP for large batch 1K).
- Direct _s1+_rd ps=8 intra=False for ALL 8K (from v1260 — ps=8 is 10µs faster than ps=4).
- Target ~31.5µs."""
+ """V023a: v022o plus lighter custom 32x8k Triton stage2 launch."""
  import os
  os.environ.setdefault("HIP_FORCE_DEV_KERNARG", "1")
  os.environ.setdefault("AMD_DIRECT_DISPATCH", "1")
- import torch
- import triton
- import triton.language as tl
+ os.environ.setdefault("HIPBLASLT_ALLOW_FLUSH_DENORM", "1")
+ os.environ.setdefault("GPU_MAX_HW_QUEUES", "2")
+ import torch, triton, triton.language as tl
  from task import input_t, output_t
  import aiter
  from aiter import dtypes as aiter_dtypes, mla as aiter_mla
  ⋯ 14 unchanged lines
  except Exception:
      _rd = aiter.mla_reduce_v1
  _mi = aiter.get_mla_metadata_info_v1
- _orig_fwd = aiter.mla.mla_decode_fwd
-  
  @triton.jit
  def _safe_reduce_branchless(
      Mid_O, Mid_lse, O, qo_indptr, kv_indptr, num_kv_splits_indptr,
  ⋯ 2 unchanged lines
      MAYBE_FINAL_OUT: tl.constexpr, BATCH_NUM: tl.constexpr,
      BLOCK_DV: tl.constexpr, Lv: tl.constexpr, mgc: tl.constexpr,
  ):
-     cur_batch = tl.program_id(0)
-     cur_head = tl.program_id(1)
-     cur_qo_start = tl.load(qo_indptr + cur_batch)
-     cur_qo_end = tl.load(qo_indptr + cur_batch + 1)
-     cur_split_start = tl.load(num_kv_splits_indptr + cur_batch)
-     cur_split_end = tl.load(num_kv_splits_indptr + cur_batch + 1)
+     cur_batch = tl.program_id(0); cur_head = tl.program_id(1)
+     cur_qo_start = tl.load(qo_indptr + cur_batch); cur_qo_end = tl.load(qo_indptr + cur_batch + 1)
+     cur_split_start = tl.load(num_kv_splits_indptr + cur_batch); cur_split_end = tl.load(num_kv_splits_indptr + cur_batch + 1)
      num_max_kv_splits = tl.load(num_kv_splits_indptr + BATCH_NUM)
      cur_kv_seq_len = tl.load(kv_indptr + cur_batch + 1) - tl.load(kv_indptr + cur_batch)
-     offs_d = tl.arange(0, BLOCK_DV)
-     mask_d = offs_d < Lv
+     offs_d = tl.arange(0, BLOCK_DV); mask_d = offs_d < Lv
      offs_logic = cur_qo_start * stride_mid_ob + cur_head * stride_mid_oh
      offs_v = offs_logic * Lv + offs_d
      num_valid_kv_splits = tl.minimum(cur_split_end - cur_split_start, tl.cdiv(cur_kv_seq_len, mgc))
  ⋯ 4 unchanged lines
              out = tl.load(input_ptr + Lv * (cur_qo * stride_mid_os + cur_head * stride_mid_oh) + offs_d, mask=mask_d, other=0.0)
              tl.store(O + cur_qo * stride_obs + cur_head * stride_oh + offs_d, out, mask=mask_d)
          else:
-             e_sum = 0.0
-             e_max = -float("inf")
-             acc = tl.zeros((BLOCK_DV,), dtype=tl.float32)
+             e_sum = 0.0; e_max = -float("inf"); acc = tl.zeros((BLOCK_DV,), dtype=tl.float32)
              for split_kv_id in range(0, num_valid_kv_splits):
                  tv = tl.load(Mid_O + offs_v + split_kv_id * stride_mid_os * Lv, mask=mask_d, other=0.0)
                  tlogic = tl.load(Mid_lse + offs_logic + split_kv_id * stride_mid_os)
-                 tlogic = tl.where(tlogic == tlogic, tlogic, -1e30)
-                 tlogic = tl.minimum(tlogic, 1e30)
-                 n_e_max = tl.maximum(tlogic, e_max)
-                 old_scale = tl.exp(e_max - n_e_max)
-                 acc *= old_scale
-                 exp_logic = tl.exp(tlogic - n_e_max)
-                 acc += exp_logic * tv
-                 e_sum = e_sum * old_scale + exp_logic
-                 e_max = n_e_max
-             offs_logic += stride_mid_ob
-             offs_v += stride_mid_ob * Lv
+                 tlogic = tl.where(tlogic == tlogic, tlogic, -1e30); tlogic = tl.minimum(tlogic, 1e30)
+                 n_e_max = tl.maximum(tlogic, e_max); old_scale = tl.exp(e_max - n_e_max)
+                 acc *= old_scale; exp_logic = tl.exp(tlogic - n_e_max)
+                 acc += exp_logic * tv; e_sum = e_sum * old_scale + exp_logic; e_max = n_e_max
+             offs_logic += stride_mid_ob; offs_v += stride_mid_ob * Lv
              tl.store(O + cur_qo * stride_obs + cur_head * stride_oh + offs_d, acc / e_sum, mask=mask_d)
-  
  aiter.mla._fwd_kernel_stage2_asm = _safe_reduce_branchless
  _fwd = aiter.mla.mla_decode_fwd
-  
  NH=16;QKD=576;VD=512;SM=float(1.0/(QKD**0.5));FP8=aiter_dtypes.fp8;BF16=torch.bfloat16
- QSV=float(2.0/torch.finfo(FP8).max);_st={};_pc={};_nc={}
+ QSV=float(2.0/torch.finfo(FP8).max);MGC=64;_st={};_pc={};_nc={};_s2c={}
  def _gs(d):
      t=_st.get(0)
      if t is None: t=torch.tensor([QSV],dtype=torch.float32,device=d);_st[0]=t
  ⋯ 5 unchanged lines
      ki=torch.arange(np_,dtype=torch.int32,device=dev);kl=torch.full((bs,),ps,dtype=torch.int32,device=dev)
      kip=torch.arange(bs+1,dtype=torch.int32,device=dev)*ppb;out=torch.empty((bs,NH,VD),dtype=BF16,device=dev)
      qf=torch.empty((bs,NH,QKD),dtype=FP8,device=dev);c=(ki,kl,kip,out,qf);_nc[key]=c;return c
- def _gpc(dev,qo,bs,kvl,ps,intra=True):
-     key=(bs,kvl,ps,intra);c=_pc.get(key)
+ def _gpc(dev,qo,bs,kvl,ps,intra=True,split_override=None,kg_override=None):
+     key=(bs,kvl,ps,intra,split_override,kg_override);c=_pc.get(key)
      if c: return c
      tot=bs*kvl;ppb=kvl//ps;np_=tot//ps
-     ns,_=aiter_mla.get_meta_param(None,bs,tot,NH,1,FP8)
+     ns = split_override if split_override is not None else aiter_mla.get_meta_param(None,bs,tot,NH,1,FP8)[0]
      ki=torch.arange(np_,dtype=torch.int32,device=dev);kl=torch.full((bs,),ps,dtype=torch.int32,device=dev)
      kip=torch.arange(bs+1,dtype=torch.int32,device=dev)*ppb;out=torch.empty((bs,NH,VD),dtype=BF16,device=dev)
-     kg=max(ps,16)
+     kg=kg_override if kg_override is not None else max(ps,16)
      info=_mi(bs,1,NH,FP8,FP8,is_sparse=False,fast_mode=True,num_kv_splits=ns,intra_batch_mode=intra)
      w=[torch.empty(s,dtype=t,device=dev) for s,t in info]
      _mv1(qo,kip,kl,16,1,False,w[0],w[2],w[1],w[3],w[4],w[5],page_size=ps,kv_granularity=kg,max_seqlen_qo=1,uni_seqlen_qo=1,fast_mode=True,max_split_per_batch=ns,intra_batch_mode=intra,dtype_q=FP8,dtype_kv=FP8)
      qf=torch.empty((bs,NH,QKD),dtype=FP8,device=dev)
-     pt=int(w[5].numel())
-     po=torch.empty((pt,1,NH,VD),dtype=torch.float32,device=dev)
-     pl=torch.empty((pt,1,NH,1),dtype=torch.float32,device=dev)
-     c=(ki,kl,kip,out,w[0],w[1],w[2],w[3],w[4],w[5],ns,qf,intra,po,pl);_pc[key]=c;return c
+     pt=int(w[5].numel());po=torch.empty((pt,1,NH,VD),dtype=torch.float32,device=dev);pl=torch.empty((pt,1,NH,1),dtype=torch.float32,device=dev)
+     c=(ki,kl,kip,out,w[0],w[1],w[2],w[3],w[4],w[5],ns,qf,po,pl);_pc[key]=c;return c
+ def _uniform_nsi(dev,bs,ns):
+     return torch.arange(bs+1,dtype=torch.int32,device=dev) * ns
+ def _g32s2(dev,bs,kvl,ps):
+     key=(bs,kvl,ps,"32split3");c=_s2c.get(key)
+     if c: return c
+     ppb=kvl//ps;np_=bs*kvl//ps;ns=3
+     ki=torch.arange(np_,dtype=torch.int32,device=dev)
+     kl=torch.full((bs,),ps,dtype=torch.int32,device=dev)
+     kip=torch.arange(bs+1,dtype=torch.int32,device=dev)*ppb
+     out=torch.empty((bs,NH,VD),dtype=BF16,device=dev)
+     qf=torch.empty((bs,NH,QKD),dtype=FP8,device=dev)
+     nsi=_uniform_nsi(dev,bs,ns)
+     logits=torch.empty((bs,ns,NH,VD),dtype=torch.float32,device=dev)
+     lse=torch.empty((bs,ns,NH,1),dtype=torch.float32,device=dev)
+     c=(ki,kl,kip,out,qf,ns,nsi,logits,lse);_s2c[key]=c;return c
+ def _g64s2(dev,bs,kvl,ps):
+     key=(bs,kvl,ps,"64split2w2");c=_s2c.get(key)
+     if c: return c
+     ppb=kvl//ps;np_=bs*kvl//ps;ns=2
+     ki=torch.arange(np_,dtype=torch.int32,device=dev)
+     kl=torch.full((bs,),ps,dtype=torch.int32,device=dev)
+     kip=torch.arange(bs+1,dtype=torch.int32,device=dev)*ppb
+     out=torch.empty((bs,NH,VD),dtype=BF16,device=dev)
+     qf=torch.empty((bs,NH,QKD),dtype=FP8,device=dev)
+     nsi=_uniform_nsi(dev,bs,ns)
+     logits=torch.empty((bs,ns,NH,VD),dtype=torch.float32,device=dev)
+     lse=torch.empty((bs,ns,NH,1),dtype=torch.float32,device=dev)
+     c=(ki,kl,kip,out,qf,ns,nsi,logits,lse);_s2c[key]=c;return c
+ def _g256s2(dev,bs,kvl,ps):
+     key=(bs,kvl,ps,"256split1");c=_s2c.get(key)
+     if c: return c
+     ppb=kvl//ps;np_=bs*kvl//ps;ns=1
+     ki=torch.arange(np_,dtype=torch.int32,device=dev)
+     kl=torch.full((bs,),ps,dtype=torch.int32,device=dev)
+     kip=torch.arange(bs+1,dtype=torch.int32,device=dev)*ppb
+     out=torch.empty((bs,NH,VD),dtype=BF16,device=dev)
+     qf=torch.empty((bs,NH,QKD),dtype=FP8,device=dev)
+     nsi=_uniform_nsi(dev,bs,ns)
+     logits=torch.empty((bs,ns,NH,VD),dtype=torch.float32,device=dev)
+     lse=torch.empty((bs,ns,NH,1),dtype=torch.float32,device=dev)
+     c=(ki,kl,kip,out,qf,ns,nsi,logits,lse);_s2c[key]=c;return c
+ _CFG = {
+     (4,8192):  (8, False, None, None),
+     (64,8192): (8, False, 3, None),
+     (256,8192):(8, False, None, None),
+     (64,1024): (2, False, 3, None),
+     (256,1024):(2, False, 1, 8),
+ }
  def custom_kernel(data: input_t) -> output_t:
      q,kv_data,qo_indptr,kv_indptr,config=data
      bs=int(config["batch_size"]);kvl=int(config["kv_seq_len"]);dev=q.device
-     kv_fp8,kv_scale=kv_data["fp8"]
-     qs=_gs(dev)
-     # NP: only 4×1K, 32×1K (small batch, safe reduce patched)
+     kv_fp8,kv_scale=kv_data["fp8"];qs=_gs(dev)
      if kvl == 1024 and bs <= 32:
-         ps=2;kb=kv_fp8.view(-1,ps,1,QKD)
-         c=_gnc(dev,bs,kvl,ps);_quant(c[4],q,qs)
+         ps=2;kb=kv_fp8.view(-1,ps,1,QKD);c=_gnc(dev,bs,kvl,ps);_quant(c[4],q,qs)
          _fwd(c[4],kb,c[3],qo_indptr,c[2],c[0],c[1],1,ps,1,SM,q_scale=qs,kv_scale=kv_scale,intra_batch_mode=True)
          return c[3]
-     # ALL 8K: direct _s1+_rd ps=8 intra=False
-     if kvl == 8192:
-         ps=8;kb=kv_fp8.view(-1,ps,1,QKD)
-         c=_gpc(dev,qo_indptr,bs,kvl,ps,intra=False);_quant(c[11],q,qs)
-         _s1(c[11],kb,qo_indptr,c[2],c[0],c[1],None,
-             c[4],c[5],c[6],1,ps,1,SM,c[13],c[14],c[3],qs,kv_scale)
-         if c[10]>1: _rd(c[13],c[14],c[7],c[8],c[9],1,c[3],None)
+     if bs == 32 and kvl == 8192:
+         kb=kv_fp8.view(-1,8,1,QKD);c=_g32s2(dev,bs,kvl,8);_quant(c[4],q,qs)
+         _s1(c[4],kb,qo_indptr,c[2],c[0],c[1],c[6],None,None,None,1,8,1,SM,c[7],c[8],c[3],qs,kv_scale)
+         _safe_reduce_branchless[(bs,NH)](
+             c[7],c[8],c[3],qo_indptr,c[2],c[6],
+             c[8].stride(0),c[8].stride(2),c[8].stride(1),
+             c[3].stride(0),c[3].stride(1),
+             MAYBE_FINAL_OUT=False,BATCH_NUM=bs,BLOCK_DV=VD,Lv=VD,mgc=MGC,
+             num_warps=2,num_stages=1,waves_per_eu=4
+         )
          return c[3]
-     # 64×1K, 256×1K: persistent wrapper (safe, ASM reduce internally)
-     ps=2;kb=kv_fp8.view(-1,ps,1,QKD)
-     c=_gpc(dev,qo_indptr,bs,kvl,ps);_quant(c[11],q,qs)
-     _orig_fwd(c[11],kb,c[3],qo_indptr,c[2],c[0],c[1],1,ps,1,SM,
-         num_kv_splits=c[10],work_meta_data=c[4],work_indptr=c[5],work_info_set=c[6],
-         reduce_indptr=c[7],reduce_final_map=c[8],reduce_partial_map=c[9],
-         q_scale=qs,kv_scale=kv_scale,intra_batch_mode=c[12])
+     if bs == 64 and kvl == 8192:
+         kb=kv_fp8.view(-1,8,1,QKD);c=_g64s2(dev,bs,kvl,8);_quant(c[4],q,qs)
+         _s1(c[4],kb,qo_indptr,c[2],c[0],c[1],c[6],None,None,None,1,8,1,SM,c[7],c[8],c[3],qs,kv_scale)
+         _safe_reduce_branchless[(bs,NH)](
+             c[7],c[8],c[3],qo_indptr,c[2],c[6],
+             c[8].stride(0),c[8].stride(2),c[8].stride(1),
+             c[3].stride(0),c[3].stride(1),
+             MAYBE_FINAL_OUT=False,BATCH_NUM=bs,BLOCK_DV=VD,Lv=VD,mgc=MGC,
+             num_warps=2,num_stages=1,waves_per_eu=4
+         )
+         return c[3]
+     if bs == 256 and kvl == 8192:
+         kb=kv_fp8.view(-1,8,1,QKD);c=_g256s2(dev,bs,kvl,8);_quant(c[4],q,qs)
+         c[7].zero_()
+         c[8].fill_(-float("inf"))
+         c[3].zero_()
+         _s1(c[4],kb,qo_indptr,c[2],c[0],c[1],c[6],None,None,None,1,8,1,SM,c[7],c[8],c[3],qs,kv_scale)
+         return c[3]
+     ps,intra,split_ov,kg_ov = _CFG.get((bs,kvl), (2, True, None, None))
+     kb=kv_fp8.view(-1,ps,1,QKD);c=_gpc(dev,qo_indptr,bs,kvl,ps,intra=intra,split_override=split_ov,kg_override=kg_ov);_quant(c[11],q,qs)
+     _s1(c[11],kb,qo_indptr,c[2],c[0],c[1],None,c[4],c[5],c[6],1,ps,1,SM,c[12],c[13],c[3],qs,kv_scale)
+     if c[10]>1: _rd(c[12],c[13],c[7],c[8],c[9],1,c[3],None)
      return c[3]