feat(op): support varlen npu flash attention (InternLM#209)

internlm/core/parallel/comm/isp.py

@@ -900,7 +900,12 @@ def auto_wrap_distributed_attention(cls: nn.Module) -> Callable[[bool, Any, floa
     def _attetion_constructor(
         local_attn_cls: type, causal=False, softmax_scale=None, attention_dropout=0.0
     ) -> nn.Module:
-        if gpc.config.parallel["tensor"].get("mode", "mtp") != "isp":
+        try:
+            tp_mode = gpc.config.parallel["tensor"].get("mode", "mtp")
+        except AttributeError:
+            tp_mode = "mtp"
+
+        if tp_mode != "isp":
             return local_attn_cls(causal, softmax_scale, attention_dropout)
         else:
             return DistributedAttention(

internlm/model/ops/attention.py

@@ -79,7 +79,7 @@ def _nyi_attn(func_name, *args, **kwargs):  # pylint: disable=W0613
 
 def _flash_float32_compatibility_wrapper(input_idxs: Tuple, flash_func: Callable, *args, **kwargs):
     if gpc.config.model.dtype is torch.float32:
-        inputs = (args[idx] for idx in input_idxs)
+        inputs = [args[idx] for idx in input_idxs]
         input_dtype = inputs[0].dtype
         other_args = [args[idx] for idx in range(len(inputs), len(args))]
 
@@ -194,10 +194,35 @@ def _flash_fixedlen_qkvsplited_attn(q, k, v, dropout_p=0.0, softmax_scale=None,
 
 
 # npu flash attention operators
-# TODO: should we add _flash_float32_compatibility_wrapper support for npu.
+def _npu_varlen_qkvsplited_attn(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens_q,
+    cu_seqlens_k,
+    max_seqlen_q,  # pylint: disable=W0613
+    max_seqlen_k,  # pylint: disable=W0613
+    dropout_p=0.0,
+    softmax_scale=None,
+    causal=False,
+):
+    return _flash_float32_compatibility_wrapper(
+        (0, 1, 2),
+        _npu_varlen_qkvsplited_func,
+        q,
+        k,
+        v,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlen_q,
+        max_seqlen_k,
+        dropout_p,
+        softmax_scale,
+        causal,
+    )
 
 
-def _npu_varlen_qkvsplited_attn(
+def _npu_varlen_qkvsplited_func(
     q: torch.Tensor,
     k: torch.Tensor,
     v: torch.Tensor,
@@ -208,17 +233,32 @@ def _npu_varlen_qkvsplited_attn(
     dropout_p=0.0,
     softmax_scale=None,
     causal=False,
+    use_fixlen=False,
 ):
-    # TODO: support npu native varlen flash attention
+    """Support Huawei Ascend's torch_npu flash attention.
+    Tested version:
+        torch: 2.1.0+cpu
+        torch_npu: 2.1.0.post3+git7c4136d
+        cann: 8.0.RC1.alpha003
+    """
     packed_length = q.size(dim=1)
+    softmax_scale = softmax_scale or 1.0 / math.sqrt(q.shape[-1])
 
-    q = unpack_qkv_before_attn(q, cu_seqlens=cu_seqlens_q)
-    k = unpack_qkv_before_attn(k, cu_seqlens=cu_seqlens_k)
-    v = unpack_qkv_before_attn(v, cu_seqlens=cu_seqlens_k)
+    if use_fixlen:
 
-    output = _npu_fixedlen_qkvsplited_attn(q, k, v, dropout_p, softmax_scale, causal)
+        q = unpack_qkv_before_attn(q, cu_seqlens=cu_seqlens_q)
+        k = unpack_qkv_before_attn(k, cu_seqlens=cu_seqlens_k)
+        v = unpack_qkv_before_attn(v, cu_seqlens=cu_seqlens_k)
 
-    return pack_output_after_attn(output, cu_seqlens_q, packed_length)
+        output = _npu_fixedlen_qkvsplited_attn(q, k, v, dropout_p, softmax_scale, causal)
+
+        output = pack_output_after_attn(output, cu_seqlens_q, packed_length)
+    else:
+        output = _npu_fused_varlen_qkvsplited_attn(
+            q, k, v, dropout_p, softmax_scale, causal, max_seqlen_q, max_seqlen_k, cu_seqlens_q, cu_seqlens_k
+        )
+
+    return output
 
 
 def _npu_fixedlen_qkvsplited_attn(
@@ -236,6 +276,7 @@ def _npu_fixedlen_qkvsplited_attn(
         q, k, v = q.squeeze(dim=2), k.squeeze(dim=2), v.squeeze(dim=2)
 
     _, seqlen, n_head, _ = q.shape
+    sparse_mode = 0
     attention_mask = torch.triu(torch.ones(seqlen, seqlen, device=get_current_device()), 1).bool()
 
     return _origin_npu_fixedlen_qkvsplited_func(
@@ -247,25 +288,71 @@ def _npu_fixedlen_qkvsplited_attn(
         pse=None,
         atten_mask=attention_mask,
         scale=softmax_scale,
-        sparse_mode=0,  # If necessary, expose the interface
+        sparse_mode=sparse_mode,  # If necessary, expose the interface
         pre_tockens=seqlen,  # Used for sparse calculations, representing the left boundary of the slides window
         next_tockens=0,  # If necessary, expose the interface
         keep_prob=1 - dropout_p,
         inner_precise=0,  # If necessary, expose the interface
-    )
+    )[0]
 
 
-def _npu_varlen_qkvpacked_attn(
-    qkv: torch.Tensor, cu_seqlens, max_seqlen, dropout_p, softmax_scale=None, causal=False  # pylint: disable=W0613
+def _npu_fused_varlen_qkvsplited_attn(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    dropout_p: float,
+    softmax_scale=None,
+    causal=False,
+    max_seqlen_q: int = None,
+    max_seqlen_k: int = None,
+    cu_seqlens_q=None,
+    cu_seqlens_kv=None,
+    deterministic=False,
 ):
-    # TODO: support npu native varlen flash attention
-    packed_length = qkv.size(dim=1)
+    assert causal is True
+    assert q.dtype in (torch.bfloat16, torch.float16)
 
-    qkv = unpack_qkv_before_attn(qkv, cu_seqlens=cu_seqlens)
+    if len(q.shape) == 4:  # [1, packedseqlen, n_head, headdim]
+        q, k, v = q.squeeze(dim=0), k.squeeze(dim=0), v.squeeze(dim=0)
 
-    output = _npu_fixedlen_qkvpacked_attn(qkv, dropout_p, softmax_scale, causal)
+    S, N = max(max_seqlen_q, max_seqlen_k), q.shape[1]
+    device = get_current_device()
+    sparse_mode = 0
 
-    return pack_output_after_attn(output, cu_seqlens, packed_length)
+    if max_seqlen_k > 2048 and max_seqlen_q > 2048:
+        sparse_mode = 2
+        max_seqlen_k = 2048
+        max_seqlen_q = 2048
+
+    attention_mask = torch.triu(torch.ones(max_seqlen_q, max_seqlen_k, device=device), 1).bool()
+    cu_seqlens_q = cu_seqlens_q[1:].tolist()
+    cu_seqlens_kv = cu_seqlens_kv[1:].tolist()
+
+    return _origin_npu_fixedlen_qkvsplited_func(
+        query=q,
+        key=k,
+        value=v,
+        head_num=N,
+        input_layout="TND",
+        pse=None,
+        atten_mask=attention_mask,
+        scale=softmax_scale,
+        sparse_mode=sparse_mode,
+        pre_tockens=S,  # Used for sparse calculations, representing the left boundary of the slides window
+        next_tockens=0,
+        keep_prob=1 - dropout_p,
+        inner_precise=0 if not deterministic else 2,
+        actual_seq_kvlen=cu_seqlens_kv,
+        actual_seq_qlen=cu_seqlens_q,
+    )[0].unsqueeze(dim=0)
+
+
+def _npu_varlen_qkvpacked_attn(
+    qkv: torch.Tensor, cu_seqlens, max_seqlen, dropout_p, softmax_scale=None, causal=False  # pylint: disable=W0613
+):
+    # TODO: support npu native varlen flash attention
+    q, k, v = qkv.unbind(dim=2)
+    return _npu_varlen_qkvsplited_attn(q, k, v, cu_seqlens, max_seqlen, dropout_p, softmax_scale, causal)
 
 
 def _npu_fixedlen_qkvpacked_attn(qkv: torch.Tensor, dropout_p: float, softmax_scale=None, causal=False):
@@ -285,14 +372,20 @@ def _npu_varlen_kvpacked_attn(
     causal=False,
 ):
     # TODO: support npu native varlen flash attention
-    packed_length = q.size(dim=1)
-
-    q = unpack_qkv_before_attn(q, cu_seqlens=cu_seqlens_q)
-    kv = unpack_qkv_before_attn(kv, cu_seqlens=cu_seqlens_k)
-
-    output = _npu_fixedlen_kvpacked_attn(q, kv, dropout_p, softmax_scale, causal)
-
-    return pack_output_after_attn(output, cu_seqlens_q, packed_length)
+    k, v = kv.unbind(dim=2)
+    k, v = k.squeeze(dim=2), v.squeeze(dim=2)
+    return _npu_varlen_qkvsplited_attn(
+        q,
+        k,
+        v,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlen_q,
+        max_seqlen_k,
+        dropout_p,
+        softmax_scale,
+        causal,
+    )
 
 
 def _npu_fixedlen_kvpacked_attn(q: torch.Tensor, kv: torch.Tensor, dropout_p: float, softmax_scale=None, causal=False):
@@ -335,12 +428,6 @@ def _deeplink_fixedlen_qkvsplited_attn(*args, **kwargs):
 
 
 # torch attention operators
-
-
-def _torch_varlen_qkvpacked_attn(*args, **kwargs):
-    _nyi_attn("_torch_varlen_qkvpacked_attn", *args, **kwargs)
-
-
 # adpated from https://github.com/Dao-AILab/flash-attention/blob/v2.2.1/flash_attn/modules/mha.py
 def _torch_fixedlen_qkvpacked_attn(qkv: torch.Tensor, dropout, softmax_scale=None, causal=False, key_padding_mask=None):
     batch_size, seqlen = qkv.shape[0], qkv.shape[1]
@@ -369,10 +456,6 @@ def _torch_fixedlen_qkvpacked_attn(qkv: torch.Tensor, dropout, softmax_scale=Non
     return output
 
 
-def _torch_varlen_kvpacked_attn(*args, **kwargs):
-    _nyi_attn("_torch_varlen_kvpacked_attn", *args, **kwargs)
-
-
 # adpated from https://github.com/Dao-AILab/flash-attention/blob/v2.2.1/flash_attn/modules/mha.py
 def _torch_fixedlen_kvpacked_attn(
     q: torch.Tensor, kv: torch.Tensor, dropout, softmax_scale=None, causal=False, key_padding_mask=None
@@ -407,17 +490,78 @@ def _torch_fixedlen_kvpacked_attn(
     return output
 
 
-def _torch_varlen_qkvsplited_attn(*args, **kwargs):
-    _nyi_attn("_torch_varlen_qkvsplited_attn", *args, **kwargs)
-
-
 def _torch_fixedlen_qkvsplited_attn(
     q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, dropout, softmax_scale=None, causal=False, key_padding_mask=None
 ):
     kv = torch.stack([k, v], dim=2)
     return _torch_fixedlen_kvpacked_attn(q, kv, dropout, softmax_scale, causal, key_padding_mask)
 
 
+def _torch_varlen_qkvsplited_attn(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens_q,
+    cu_seqlens_k,
+    max_seqlen_q,  # pylint: disable=W0613
+    max_seqlen_k,  # pylint: disable=W0613
+    dropout,
+    softmax_scale=None,
+    causal=False,
+    key_padding_mask=None,
+):
+    kv = torch.stack([k, v], dim=2)
+    packed_length = q.size(dim=1)
+
+    q = unpack_qkv_before_attn(q, cu_seqlens=cu_seqlens_q)
+    kv = unpack_qkv_before_attn(kv, cu_seqlens=cu_seqlens_k)
+
+    output = _torch_fixedlen_kvpacked_attn(q, kv, dropout, softmax_scale, causal, key_padding_mask)
+
+    return pack_output_after_attn(output, cu_seqlens_q, packed_length)
+
+
+def _torch_varlen_qkvpacked_attn(
+    qkv: torch.Tensor,
+    cu_seqlens,
+    max_seqlen,  # pylint: disable=W0613
+    dropout,
+    softmax_scale=None,
+    causal=False,
+    key_padding_mask=None,
+):
+
+    packed_length = qkv.size(dim=1)
+    qkv = unpack_qkv_before_attn(qkv, cu_seqlens=cu_seqlens)
+
+    output = _torch_fixedlen_qkvpacked_attn(qkv, dropout, softmax_scale, causal, key_padding_mask)
+
+    return pack_output_after_attn(output, cu_seqlens, packed_length)
+
+
+def _torch_varlen_kvpacked_attn(
+    q: torch.Tensor,
+    kv: torch.Tensor,
+    cu_seqlens_q,
+    cu_seqlens_k,
+    max_seqlen_q,  # pylint: disable=W0613
+    max_seqlen_k,  # pylint: disable=W0613
+    dropout,
+    softmax_scale=None,
+    causal=False,
+    key_padding_mask=None,
+):
+
+    packed_length = q.size(dim=1)
+
+    q = unpack_qkv_before_attn(q, cu_seqlens=cu_seqlens_q)
+    kv = unpack_qkv_before_attn(kv, cu_seqlens=cu_seqlens_k)
+
+    output = _torch_fixedlen_kvpacked_attn(q, kv, dropout, softmax_scale, causal, key_padding_mask)
+
+    return pack_output_after_attn(output, cu_seqlens_q, packed_length)
+
+
 @auto_wrap_distributed_attention
 class SelfAttention(nn.Module):
     """Implements scaled dot-product attention with optional softmax scaling.