Use triton fa by default

vllm/attention/backends/rocm_flash_attn.py

@@ -229,9 +229,11 @@ class ROCmFlashAttentionImpl(AttentionImpl):
         # NOTE: Allow for switching between Triton and CK. Defaulting to triton.
         self.use_triton_flash_attn = envs.VLLM_USE_TRITON_FLASH_ATTN
         if self.use_triton_flash_attn:
-            from vllm.attention.ops.triton_flash_attention import (  # noqa: F401
-                triton_attention)
-            self.attn_func = triton_attention
+            # from vllm.attention.ops.triton_flash_attention import (  # noqa: F401
+            #     triton_attention)
+            from vllm.attention.ops.flash_attn_triton_mqa_gqa import ( 
+                flash_attn_varlen_func)
+            self.attn_func = flash_attn_varlen_func # triton_attention
             logger.debug("Using Triton FA in ROCmBackend")
         else:
             # if not using triton, navi3x/navi21/navi10 do not use flash-attn
@@ -325,17 +327,27 @@ class ROCmFlashAttentionImpl(AttentionImpl):
                 # When block_tables are not filled, it means q and k are the
                 # prompt, and they have the same length.
                 if self.use_triton_flash_attn:
-                    out, _ = self.attn_func(
-                        query,
-                        key,
-                        value,
-                        None,
-                        prefill_meta.seq_start_loc,
-                        prefill_meta.seq_start_loc,
-                        prefill_meta.max_prefill_seq_len,
-                        prefill_meta.max_prefill_seq_len,
-                        True,
-                        self.scale,
+                    # out, _ = self.attn_func(
+                    #     query,
+                    #     key,
+                    #     value,
+                    #     None,
+                    #     prefill_meta.seq_start_loc,
+                    #     prefill_meta.seq_start_loc,
+                    #     prefill_meta.max_prefill_seq_len,
+                    #     prefill_meta.max_prefill_seq_len,
+                    #     True,
+                    #     self.scale,
+                    out = self.attn_func(
+                        q=query,
+                        k=key,
+                        v=value,
+                        cu_seqlens_q=prefill_meta.seq_start_loc,
+                        cu_seqlens_k=prefill_meta.seq_start_loc,
+                        max_seqlens_q=prefill_meta.max_prefill_seq_len,
+                        max_seqlens_k=prefill_meta.max_prefill_seq_len,
+                        softmax_scale=self.scale,
+                        causal=True,
                     )
                 elif self.use_naive_attn:
                     if self.num_kv_heads != self.num_heads:

vllm/attention/ops/flash_attn_triton_mqa_gqa.py

+#!/usr/bin/env python
+"""
+Fused Attention
+===============
+
+This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao (https://tridao.me/publications/flash2/flash2.pdf)
+Credits: OpenAI kernel team, AMD ML Frameworks Triton team
+
+Features supported:
+
+1) Fwd with causal masking
+2) Any sequence lengths without padding (currently fwd kernel only)
+3) Support for different sequence lengths for q and k
+4) Nested tensor API currently does not support dropout or bias.
+
+Not currently supported:
+
+1) Non power of two head dims
+
+"""
+
+import argparse
+import pytest
+import random
+import sys
+import torch
+
+import triton
+import triton.language as tl
+
+torch_dtype:tl.constexpr = torch.float16
+
+TORCH_HAS_FP8E5 = hasattr(torch, 'float8_e5m2fnuz')
+if TORCH_HAS_FP8E5:
+    torch_dtype:tl.constexpr = torch.float8_e5m2fnuz
+
+class MetaData():
+    cu_seqlens_q = None
+    cu_seqlens_k = None
+    max_seqlens_q = 0
+    max_seqlens_k = 0
+    bias = None
+    alibi_slopes = None
+    causal = False
+    num_contexts = 0
+    varlen = False
+    dropout_p, return_encoded_softmax = 0.0, False
+
+    def __init__(self, sm_scale=1.0, causal=False, dropout_p=0.0, return_encoded_softmax=False):
+        self.sm_scale = sm_scale
+        self.causal = causal
+        self.dropout_p = dropout_p
+        self.return_encoded_softmax = return_encoded_softmax
+
+    def set_varlen_params(self, cu_seqlens_q, cu_seqlens_k):
+        self.varlen = True
+        self.cu_seqlens_q = cu_seqlens_q
+        self.cu_seqlens_k = cu_seqlens_k
+        # Without "varlen", there should still be one sequence.
+        assert len(cu_seqlens_q) >= 2
+        assert len(cu_seqlens_q) == len(cu_seqlens_k)
+        self.num_contexts = len(cu_seqlens_q) - 1
+        for i in range (0, self.num_contexts):
+            self.max_seqlens_q = max(cu_seqlens_q[i+1].item() - cu_seqlens_q[i].item(), self.max_seqlens_q)
+            self.max_seqlens_k = max(cu_seqlens_k[i+1].item() - cu_seqlens_k[i].item(), self.max_seqlens_k)
+
+    def need_bias(self, bias, batch, nheads, seqlen_q, seqlen_k):
+        assert bias.is_cuda
+        assert bias.dim() == 4
+        assert bias.shape[0] == 1
+        assert bias.shape[2:] == (seqlen_q, seqlen_k)
+        self.bias = bias
+
+    def need_alibi(self, alibi_slopes, batch, nheads):
+        assert alibi_slopes.is_cuda
+        assert alibi_slopes.dim() == 2
+        assert alibi_slopes.shape[0] == batch
+        assert alibi_slopes.shape[1] == nheads
+        self.alibi_slopes = alibi_slopes
+
+    def need_causal(self):
+        self.causal = True
+
+    def need_dropout(dropout_p, return_encoded_softmax):
+        self.dropout_p = dropout_p
+        self.return_encoded_softmax = return_encoded_softmax
+
+    def check_args(self, q, k, v, o):
+        assert q.dim() == k.dim() and q.dim() == v.dim()
+        if self.varlen:
+            assert q.dim() == 3
+            total_q, nheads_q, head_size = q.shape
+            total_k, nheads_k, _ = k.shape
+            assert self.cu_seqlens_q is not None
+            assert self.cu_seqlens_k is not None
+            assert len(self.cu_seqlens_q) == len(self.cu_seqlens_k)
+            # TODO: Remove once bias is supported with varlen
+            assert self.bias == None
+            # TODO:Remove once dropout is supported with varlen
+            assert self.dropout_p == 0.0
+            assert not self.return_encoded_softmax
+        else:
+            assert q.dim() == 4
+            batch, nheads_q, seqlen_q, head_size = q.shape
+            _, nheads_k, seqlen_k, _ = k.shape
+            assert self.max_seqlens_q > 0 and self.max_seqlens_k > 0
+            assert self.cu_seqlens_q is None and self.cu_seqlens_k is None
+        assert k.shape == v.shape
+        assert q.shape[-1] == k.shape[-1] and q.shape[-1] == v.shape[-1]
+        # TODO: Change assert if we support qkl f8 and v f16
+        assert q.dtype == k.dtype and q.dtype == v.dtype
+        assert head_size <= 256
+        assert o.shape == q.shape
+        assert (nheads_q % nheads_k) == 0
+
+@triton.jit
+def cdiv_fn(x,y):
+    return (x + y - 1) // y
+
+@triton.jit
+def max_fn(x, y):
+    return tl.math.max(x, y)
+
+@triton.jit
+def dropout_offsets(philox_seed, philox_offset, dropout_p, m, n, stride):
+    ms = tl.arange(0, m)
+    ns = tl.arange(0, n)
+    return philox_offset + ms[:, None] * stride + ns[None, :]
+
+@triton.jit
+def dropout_rng(philox_seed, philox_offset, dropout_p, m, n, stride):
+    rng_offsets = dropout_offsets(philox_seed, philox_offset, dropout_p, m, n, stride).to(tl.uint32)
+    # TODO: use tl.randint for better performance
+    return tl.rand(philox_seed, rng_offsets)
+
+@triton.jit
+def dropout_mask(philox_seed, philox_offset, dropout_p, m, n, stride):
+    rng_output = dropout_rng(philox_seed, philox_offset, dropout_p, m, n, stride)
+    rng_keep = rng_output > dropout_p
+    return rng_keep
+
+@triton.jit
+def load_fn(block_ptr, first, second, pad):
+    if first and second:
+        tensor = tl.load(block_ptr, boundary_check=(0,1), padding_option=pad)
+    elif first:
+        tensor = tl.load(block_ptr, boundary_check=(0,), padding_option=pad)
+    elif second:
+        tensor = tl.load(block_ptr, boundary_check=(1,), padding_option=pad)
+    else:
+        tensor = tl.load(block_ptr)
+    return tensor
+
+@triton.jit
+def print_gpu(prefix, val=None):
+    if (tl.program_id(0) == 0) and ((tl.program_id(1) == 0) and (tl.program_id(2) == 0)):
+        if val is not None:
+            tl.device_print(prefix, val)
+        else:
+            tl.device_print(prefix)
+
+@triton.jit
+def compute_alibi_block(alibi_slope, seqlen_q, seqlen_k, offs_m, offs_n, transpose = False):
+    # when seqlen_k and seqlen_q are different we want the diagonal to stick to the bottom right of the attention matrix
+    # for casual mask we want something like this where (1 is kept and 0 is masked)
+    # seqlen_q = 2 and seqlen_k = 5
+    #   1 1 1 1 0
+    #   1 1 1 1 1
+    # seqlen_q = 5 and seqlen_k = 2
+    #        0 0
+    #        0 0
+    #        0 0
+    #        1 0
+    #        1 1
+    # for alibi the diagonal is 0 indicating no penalty for attending to that spot and increasing penalty for attending further from the diagonal
+    # e.g. alibi_slope = 1, seqlen_q = 2, seqlen_k = 5, offs_m = [0, 1, 2, 3], offs_n = [0, 1, 2, 3, 4], transpose = False
+    # 1. offs_m[:,None] = [[0],
+    #                       [1],
+    # 2. offs_m[:,None] + seqlen_k = [[5],
+    #                                  [6],
+    # 3. offs_m[:,None] + seqlen_k - seqlen_q = [[3],
+    #                                             [4],
+    # 4. offs_m[:,None] + seqlen_k - seqlen_q - offs_n[None,:] = [[3], - [[0, 1, 2, 3, 4]] =  [[ 3, 2, 1, 0,-1],
+    #                                                            [4],                           [ 4, 3, 2, 1, 0]]
+    # 5. -1 * alibi_slope * tl.abs(relative_pos_block) = [[ -3, -2, -1, 0,-1],
+    #                                                     [ -4, -3, -2, -1, 0]],
+    relative_pos_block = offs_m[:,None] + seqlen_k - seqlen_q - offs_n[None,:]
+    alibi_block = -1 * alibi_slope * tl.abs(relative_pos_block)
+    if transpose:
+        return alibi_block.T
+    else:
+        return alibi_block
+
+@triton.jit
+def _attn_fwd_inner(
+    acc, l_i, m_i, q,
+    K_block_ptr, V_block_ptr,
+    start_m,
+    actual_seqlen_k,
+    actual_seqlen_q,
+    dropout_p,
+    philox_seed,
+    batch_philox_offset,
+    encoded_softmax_block_ptr,
+    block_min, block_max,
+    offs_n_causal,
+    masked_blocks,
+    n_extra_tokens,
+    bias_ptr,
+    alibi_slope,
+    IS_CAUSAL: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    OFFS_M: tl.constexpr,
+    OFFS_N: tl.constexpr,
+    PRE_LOAD_V: tl.constexpr,
+    MASK_STEPS: tl.constexpr,
+    ENABLE_DROPOUT: tl.constexpr,
+    RETURN_ENCODED_SOFTMAX: tl.constexpr,
+    PADDED_HEAD: tl.constexpr
+):
+    # loop over k, v, and update accumulator
+    for start_n in range (block_min, block_max, BLOCK_N):
+        # For padded blocks, we will overrun the tensor size if
+        # we load all BLOCK_N. For others, the blocks are all within range.
+        k = load_fn(K_block_ptr, PADDED_HEAD, MASK_STEPS and (n_extra_tokens != 0), "zero")
+        if PRE_LOAD_V:
+            v = load_fn(V_block_ptr, MASK_STEPS and (n_extra_tokens != 0), PADDED_HEAD, "zero")
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        # We start from end of seqlen_k so only the first iteration would need
+        # to be checked for padding if it is not a multiple of block_n
+        # TODO: This can be optimized to only be true for the padded block.
+        if MASK_STEPS:
+            # If this is the last block / iteration, we want to
+            # mask if the sequence length is not a multiple of block size
+            # a solution is to always do BLOCK_M // BLOCK_N + 1 steps if not is_modulo_mn.
+            # last step might get wasted but that is okay. check if this masking works For
+            # that case.
+            if (start_n + BLOCK_N == block_max) and (n_extra_tokens != 0):
+                boundary_m = tl.full([BLOCK_M], actual_seqlen_k, dtype=tl.int32)
+                size_n = start_n + OFFS_N[None,:]
+                mask = size_n < boundary_m[:,None]
+                qk = tl.where(mask, qk, float("-inf"))
+        if IS_CAUSAL:
+            causal_boundary = start_n + offs_n_causal
+            causal_mask = OFFS_M[:, None] >= causal_boundary[None, :]
+            qk = tl.where(causal_mask, qk, float("-inf"))
+        # -- compute qk ----
+        qk += tl.dot(q, k)
+        if bias_ptr is not None:
+            bias = load_fn(bias_ptr, False, MASK_STEPS and (n_extra_tokens != 0), "zero")
+            # While bias is added after multiplying qk with sm_scale,
+            # our optimization to use 2^x instead of e^x results in an additional
+            # scale factor of log2(e) which we must also multiply the bias with.
+            qk += (bias * 1.44269504089)
+           
+        if alibi_slope is not None:
+            # Compute the global position of each token within the sequence
+            global_m_positions = start_m*BLOCK_M + tl.arange(0, BLOCK_M)
+            global_n_positions = start_n + tl.arange(0, BLOCK_N)
+
+            alibi_block = compute_alibi_block(alibi_slope, actual_seqlen_q, actual_seqlen_k, global_m_positions, global_n_positions)
+
+            qk += (alibi_block * 1.44269504089) # scale factor of log2(e)
+
+        # softmax
+        m_ij = tl.maximum(m_i, tl.max(qk,1))
+        qk = qk - m_ij[:, None]
+        p = tl.math.exp2(qk)
+
+        # CAVEAT: Must update l_ij before applying dropout
+        l_ij = tl.sum(p, 1)
+        if ENABLE_DROPOUT:
+            philox_offset = batch_philox_offset + start_m * BLOCK_M * actual_seqlen_k + start_n - BLOCK_N
+            keep = dropout_mask(philox_seed, philox_offset, dropout_p, BLOCK_M, BLOCK_N, actual_seqlen_k)
+            if RETURN_ENCODED_SOFTMAX:
+                tl.store(encoded_softmax_block_ptr, tl.where(keep, p, -p).to(encoded_softmax_block_ptr.type.element_ty))
+            p = tl.where(keep, p, 0.0)
+        elif RETURN_ENCODED_SOFTMAX:
+            tl.store(encoded_softmax_block_ptr, p.to(encoded_softmax_block_ptr.type.element_ty))
+        # -- update output accumulator --
+        alpha = tl.math.exp2(m_i - m_ij)
+        acc = acc * alpha[:, None]
+        if not PRE_LOAD_V:
+            v = load_fn(V_block_ptr, MASK_STEPS and (n_extra_tokens != 0), PADDED_HEAD, "zero")
+        # -- update m_i and l_i
+        l_i = l_i * alpha + l_ij
+        # update m_i and l_i
+        m_i = m_ij
+        acc += tl.dot(p.to(V_block_ptr.type.element_ty), v)
+        V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
+        K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
+        if bias_ptr is not None:
+            bias_ptr = tl.advance(bias_ptr, (0, BLOCK_N))
+        if RETURN_ENCODED_SOFTMAX:
+            encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr, (0, BLOCK_N))
+    return acc, l_i, m_i
+
+@triton.autotune(
+   configs=[
+       triton.Config({'BLOCK_M': 256, 'BLOCK_N': 64, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=8),
+       triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
+       triton.Config({'BLOCK_M': 256, 'BLOCK_N': 128, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=8),
+       triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'waves_per_eu': 0, 'PRE_LOAD_V': True}, num_stages=1, num_warps=4),
+       triton.Config({'BLOCK_M': 128, 'BLOCK_N': 64, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
+       triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'waves_per_eu': 0, 'PRE_LOAD_V': True}, num_stages=2, num_warps=4),
+       triton.Config({'BLOCK_M': 128, 'BLOCK_N': 32, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=2, num_warps=4),
+       triton.Config({'BLOCK_M': 64, 'BLOCK_N': 64, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=8),
+       triton.Config({'BLOCK_M': 64, 'BLOCK_N': 64, 'waves_per_eu': 0, 'PRE_LOAD_V': True}, num_stages=1, num_warps=4),
+       triton.Config({'BLOCK_M': 32, 'BLOCK_N': 32, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=8),
+       # TODO: This config fails with head_size not pow2 with data mismatches. Check why.
+    #    triton.Config({'BLOCK_M': 32, 'BLOCK_N': 16, 'waves_per_eu': 1, 'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
+       triton.Config({'BLOCK_M': 16, 'BLOCK_N': 16, 'waves_per_eu': 0, 'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
+   ],
+   key=['IS_CAUSAL', 'dropout_p', 'BLOCK_DMODEL'],
+#    use_cuda_graph=True,
+)
+@triton.jit
+def attn_fwd(
+    Q, K, V, bias, sm_scale, L, Out,
+    stride_qz, stride_qh, stride_qm, stride_qk,
+    stride_kz, stride_kh, stride_kn, stride_kk,
+    stride_vz, stride_vh, stride_vk, stride_vn,
+    stride_oz, stride_oh, stride_om, stride_on,
+    stride_bz, stride_bh, stride_bm, stride_bn,
+    stride_az, stride_ah,
+    cu_seqlens_q, cu_seqlens_k,
+    dropout_p, philox_seed, philox_offset_base, encoded_softmax,
+    alibi_slopes,
+    HQ: tl.constexpr, HK:tl.constexpr,
+    ACTUAL_BLOCK_DMODEL:tl.constexpr,
+    MAX_SEQLENS_Q:tl.constexpr, MAX_SEQLENS_K:tl.constexpr,
+    VARLEN: tl.constexpr,
+    IS_CAUSAL: tl.constexpr,
+    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr, BLOCK_N: tl.constexpr,
+    PRE_LOAD_V: tl.constexpr,
+    BIAS_TYPE: tl.constexpr,
+    ENABLE_DROPOUT: tl.constexpr,
+    RETURN_ENCODED_SOFTMAX: tl.constexpr,
+    USE_ALIBI: tl.constexpr,
+    BATCH_SIZE: tl.constexpr,
+):
+    start_m = tl.program_id(0)
+    off_h_q = tl.program_id(1)
+    off_z = tl.program_id(2)
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_n = tl.arange(0, BLOCK_N)
+    if VARLEN:
+        cu_seqlens_q_start = tl.load(cu_seqlens_q + off_z)
+        cu_seqlens_q_end = tl.load(cu_seqlens_q + off_z + 1)
+        seqlen_q = cu_seqlens_q_end - cu_seqlens_q_start
+        # We have a one-size-fits-all grid in id(0). Some seqlens might be too
+        # small for all start_m so for those we return early.
+        if start_m * BLOCK_M > seqlen_q:
+            return
+        cu_seqlens_k_start = tl.load(cu_seqlens_k + off_z)
+        cu_seqlens_k_end = tl.load(cu_seqlens_k + off_z + 1)
+        seqlen_k = cu_seqlens_k_end - cu_seqlens_k_start
+    else:
+        cu_seqlens_q_start = 0
+        cu_seqlens_k_start = 0
+        seqlen_q = MAX_SEQLENS_Q
+        seqlen_k = MAX_SEQLENS_K
+
+    # Now we compute whether we need to exit early due to causal masking.
+    # This is because for seqlen_q > seqlen_k, M rows of the attn scores
+    # are completely masked, resulting in 0s written to the output, and
+    # inf written to LSE. We don't need to do any GEMMs in this case.
+    # This block of code determines what N is, and if this WG is operating
+    # on those M rows.
+    n_blocks = cdiv_fn(seqlen_k, BLOCK_N)
+    if (IS_CAUSAL):
+        # If seqlen_q == seqlen_k, the attn scores are a square matrix.
+        # If seqlen_q != seqlen_k, attn scores are rectangular which means
+        # the causal mask boundary is bottom right aligned, and ends at either
+        # the top edge (seqlen_q < seqlen_k) or left edge.
+        # This captures the decrease in n_blocks if we have a rectangular attn matrix
+        n_blocks_seqlen = cdiv_fn(
+            (start_m + 1) * BLOCK_M + seqlen_k - seqlen_q,
+            BLOCK_N
+        )
+        # This is what adjusts the block_max for the current WG, only
+        # if IS_CAUSAL. Otherwise we want to always iterate through all n_blocks
+        n_blocks = min(n_blocks, n_blocks_seqlen)
+        # If we have no blocks after adjusting for seqlen deltas, this WG is part of
+        # the blocks that are all 0. We exit early.
+        if n_blocks <= 0:
+            o_offset = off_z * stride_oz + cu_seqlens_q_start * stride_om + off_h_q * stride_oh
+            O_block_ptr = tl.make_block_ptr(
+                base=Out + o_offset,
+                shape=(seqlen_q, BLOCK_DMODEL),
+                strides=(stride_om, stride_on),
+                offsets=(start_m * BLOCK_M, 0),
+                block_shape=(BLOCK_M, BLOCK_DMODEL),
+                order=(1, 0)
+            )
+            acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=Out.type.element_ty)
+            # We still need to write 0s to the result
+            tl.store(O_block_ptr, acc.to(Out.type.element_ty), boundary_check=(0,1))
+            l_ptrs = L + off_z * HQ * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q + offs_m
+            # We store inf to LSE, not -inf because in the bwd pass, we subtract this
+            # from qk which makes it -inf, such that exp(qk - inf) = 0 for these masked blocks.
+            l = tl.full([BLOCK_M], value=float("inf"), dtype=tl.float32)
+            tl.store(l_ptrs, l)
+            # TODO: Should dropout and return encoded softmax be handled here too?
+            return
+
+    # If MQA / GQA, set the K and V head offsets appropriately.
+    GROUP_SIZE: tl.constexpr = HQ // HK
+    if GROUP_SIZE != 1:
+        off_h_k = off_h_q // GROUP_SIZE
+    else:
+        off_h_k = off_h_q
+
+    need_padding = False
+    n_extra_tokens = 0
+    if seqlen_k < BLOCK_N:
+        need_padding = True
+        n_extra_tokens = BLOCK_N - seqlen_k
+    elif seqlen_k % BLOCK_N:
+        need_padding = True
+        n_extra_tokens = seqlen_k % BLOCK_N
+    PADDED_HEAD:tl.constexpr = (ACTUAL_BLOCK_DMODEL != BLOCK_DMODEL)
+
+    # Compute pointers for all the tensors used in this kernel.
+    q_offset = off_z * stride_qz +  off_h_q * stride_qh + cu_seqlens_q_start * stride_qm
+    Q_block_ptr = tl.make_block_ptr(
+        base=Q + q_offset,
+        shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_qm, stride_qk),
+        offsets=(start_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_DMODEL),
+        order=(1, 0)
+    )
+    k_offset = off_z * stride_kz + off_h_k * stride_kh + cu_seqlens_k_start * stride_kn
+    K_block_ptr = tl.make_block_ptr(
+        base=K + k_offset,
+        shape=(ACTUAL_BLOCK_DMODEL, seqlen_k),
+        strides=(stride_kk, stride_kn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_DMODEL, BLOCK_N),
+        order=(0, 1)
+    )
+    v_offset = off_z * stride_vz + off_h_k * stride_vh + cu_seqlens_k_start * stride_vk
+    V_block_ptr = tl.make_block_ptr(
+        base=V + v_offset,
+        shape=(seqlen_k, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_vk, stride_vn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_N, BLOCK_DMODEL),
+        order=(1, 0)
+    )
+    if BIAS_TYPE != 0:
+        b_offset = off_h_q * stride_bh # Note: this might get large enough to overflow on some configs
+        bias_ptr = tl.make_block_ptr(
+            base=bias + b_offset,
+            shape=(seqlen_q, seqlen_k),
+            strides=(stride_bm, stride_bn),
+            offsets=(start_m * BLOCK_M, 0),
+            block_shape=(BLOCK_M, BLOCK_N),
+            order=(1, 0),
+        )
+    else:
+        bias_ptr = None
+
+    if USE_ALIBI:
+        a_offset = off_z * stride_az +  off_h_q * stride_ah 
+        alibi_slope = tl.load(alibi_slopes + a_offset)
+    else:
+        alibi_slope = None
+
+    if ENABLE_DROPOUT:
+        batch_philox_offset = philox_offset_base + off_hz * seqlen_q * seqlen_k
+    else:
+        batch_philox_offset = 0
+    # We can ask to return the dropout mask without actually doing any dropout. In
+    # this case, we return an invalid pointer so indicate the mask is not valid.
+    # TODO: Fix encoded softmax. It currently uses just h_q in the base offset.
+    if RETURN_ENCODED_SOFTMAX:
+        encoded_softmax_block_ptr = tl.make_block_ptr(
+                base=encoded_softmax + off_h_q * seqlen_q * seqlen_k,
+                shape=(seqlen_q, seqlen_k),
+                strides=(seqlen_k, 1),
+                offsets=(start_m * BLOCK_M, 0),
+                block_shape=(BLOCK_M, BLOCK_N),
+                order=(1, 0)
+                )
+    else:
+        encoded_softmax_block_ptr = 0
+    # initialize pointer to m and l
+    m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    l_i = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+    # scale sm_scale by log_2(e) and use 2^x in the loop as we do not
+    # have native e^x support in HW.
+    qk_scale = sm_scale * 1.44269504089
+    # Q is loaded once at the beginning and shared by all N blocks.
+    q = load_fn(Q_block_ptr, True, PADDED_HEAD, "zero")
+    q = (q * qk_scale).to(Q_block_ptr.type.element_ty)
+
+    # Here we compute how many full and masked blocks we have.
+    padded_block_k = n_extra_tokens != 0
+    is_modulo_mn = not padded_block_k and (seqlen_q % BLOCK_M == 0)
+    if IS_CAUSAL:
+        # There are always at least BLOCK_M // BLOCK_N masked blocks.
+        # Additionally there might be one more due to dissimilar seqlens.
+        masked_blocks = BLOCK_M // BLOCK_N + (not is_modulo_mn)
+    else:
+        # Padding on Q does not need to be masked in the FA loop.
+        masked_blocks = padded_block_k
+    # if IS_CAUSAL, not is_modulo_mn does not always result in an additional block.
+    # In this case we might exceed n_blocks so pick the min.
+    masked_blocks = min(masked_blocks, n_blocks)
+    n_full_blocks = n_blocks - masked_blocks
+    block_min = 0
+    block_max = n_blocks * BLOCK_N
+    # Compute for full blocks. Here we set causal to false regardless of its actual
+    # value because there is no masking. Similarly we do not need padding.
+    if n_full_blocks > 0:
+        block_max = (n_blocks - masked_blocks) * BLOCK_N
+        acc, l_i, m_i = _attn_fwd_inner(
+            acc, l_i, m_i, q, K_block_ptr, V_block_ptr,
+            start_m, seqlen_k, seqlen_q,
+            dropout_p, philox_seed, batch_philox_offset, encoded_softmax_block_ptr,
+            # _, _, offs_n_causal, masked_blocks, n_extra_tokens, _
+            block_min, block_max, 0, 0, 0, bias_ptr, alibi_slope,
+            # IS_CAUSAL, ....
+            False, BLOCK_M, BLOCK_DMODEL, BLOCK_N, offs_m, offs_n,
+            # _, MASK_STEPS, ...
+            PRE_LOAD_V, False, ENABLE_DROPOUT, RETURN_ENCODED_SOFTMAX, PADDED_HEAD
+        )
+        block_min = block_max
+        block_max = n_blocks * BLOCK_N
+
+    tl.debug_barrier()
+    # Remaining blocks, if any, are full / not masked.
+    if (masked_blocks > 0):
+        if IS_CAUSAL:
+            offs_n_causal = offs_n + (seqlen_q - seqlen_k)
+        else:
+            offs_n_causal = 0
+        K_block_ptr = tl.advance(K_block_ptr, (0, n_full_blocks*BLOCK_N))
+        V_block_ptr = tl.advance(V_block_ptr, (n_full_blocks*BLOCK_N, 0))
+        if bias_ptr is not None:
+            bias_ptr = tl.advance(bias_ptr, (0, n_full_blocks*BLOCK_N))
+        if RETURN_ENCODED_SOFTMAX:
+            encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
+                                                   (0, n_full_blocks))
+        acc, l_i, m_i = _attn_fwd_inner(
+            acc, l_i, m_i, q, K_block_ptr, V_block_ptr,
+            start_m, seqlen_k, seqlen_q,
+            dropout_p, philox_seed, batch_philox_offset, encoded_softmax_block_ptr,
+            block_min, block_max, offs_n_causal, masked_blocks, n_extra_tokens, bias_ptr, alibi_slope,
+            IS_CAUSAL, BLOCK_M, BLOCK_DMODEL, BLOCK_N, offs_m, offs_n,
+            # _, MASK_STEPS, ...
+            PRE_LOAD_V, True, ENABLE_DROPOUT, RETURN_ENCODED_SOFTMAX, PADDED_HEAD
+        )
+    # epilogue
+    acc = acc / l_i[:, None]
+    if ENABLE_DROPOUT:
+        acc = acc / (1 - dropout_p)
+    # If seqlen_q > seqlen_k but the delta is not a multiple of BLOCK_M,
+    # then we have one block with a row of all NaNs which come from computing
+    # softmax over a row of all -infs (-inf - inf = NaN). We check for that here
+    # and store 0s where there are NaNs as these rows should've been zeroed out.
+    end_m_idx = (start_m + 1) * BLOCK_M
+    start_m_idx = start_m * BLOCK_M
+    causal_start_idx = seqlen_q - seqlen_k
+    acc = acc.to(Out.type.element_ty)
+    if IS_CAUSAL:
+        if causal_start_idx > start_m_idx and causal_start_idx < end_m_idx:
+            out_mask_boundary = tl.full((BLOCK_DMODEL,), causal_start_idx, dtype=tl.int32)
+            mask_m_offsets = start_m_idx + tl.arange(0, BLOCK_M)
+            out_ptrs_mask = mask_m_offsets[:, None] >= out_mask_boundary[None, :]
+            z = 0.0
+            acc = tl.where(out_ptrs_mask, acc, z.to(acc.type.element_ty))
+    # write back LSE
+    l_ptrs = L + off_z * HQ * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q + offs_m
+    # If seqlen_q not multiple of BLOCK_M, we need to mask out the last few rows.
+    # This is only true for the last M block. For others, overflow_size will be -ve
+    overflow_size = end_m_idx - seqlen_q
+    if overflow_size > 0:
+        boundary = tl.full((BLOCK_M,), BLOCK_M - overflow_size, dtype=tl.int32)
+        # This is a > check because mask being 0 blocks the store.
+        l_ptrs_mask = boundary > tl.arange(0, BLOCK_M)
+        tl.store(l_ptrs, m_i + tl.math.log2(l_i), mask=l_ptrs_mask)
+    else:
+        tl.store(l_ptrs, m_i + tl.math.log2(l_i))
+
+    # write back O
+    o_offset = off_z * stride_oz + cu_seqlens_q_start * stride_om + off_h_q * stride_oh
+    O_block_ptr = tl.make_block_ptr(
+        base=Out + o_offset,
+        shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_om, stride_on),
+        offsets=(start_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_DMODEL),
+        order=(1, 0)
+    )
+    # Need boundary check on this to make sure the padding from the
+    # Q and KV tensors in both dims are not part of what we store back.
+    # TODO: Do the boundary check optionally.
+    tl.store(O_block_ptr, acc, boundary_check=(0,1))
+
+@triton.jit
+def _attn_bwd_preprocess(
+    Out, DO,
+    Delta,
+    stride_oz, stride_oh, stride_om, stride_on,
+    stride_doz, stride_doh, stride_dom, stride_don,
+    seqlen_q,
+    head_dim,
+    BLOCK_M: tl.constexpr,
+    D_HEAD: tl.constexpr,
+):
+    # off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
+    # off_n = tl.arange(0, D_HEAD)
+    off_m = tl.program_id(0) * BLOCK_M
+    off_h = tl.program_id(1) # head index
+    off_z = tl.program_id(2) # batch index
+    num_h = tl.num_programs(1)
+    o_offset = off_h * stride_oh + off_z * stride_oz
+    O_block_ptr = tl.make_block_ptr(
+        base=Out + o_offset,
+        shape=(seqlen_q, head_dim),
+        strides=(stride_om, stride_on),
+        offsets=(off_m, 0),
+        block_shape=(BLOCK_M, D_HEAD),
+        order=(1, 0)
+    )
+    do_offset = off_h * stride_doh + off_z * stride_doz
+    DO_block_ptr = tl.make_block_ptr(
+        base=DO + do_offset,
+        shape=(seqlen_q, head_dim),
+        strides=(stride_dom, stride_don),
+        offsets=(off_m, 0),
+        block_shape=(BLOCK_M, D_HEAD),
+        order=(1, 0)
+    )
+    # load
+    # o = tl.load(Out + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
+    # do = tl.load(DO + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
+    o = tl.load(O_block_ptr, boundary_check=(0,1), padding_option="zero").to(tl.float32)
+    do = tl.load(DO_block_ptr, boundary_check=(0,1), padding_option="zero").to(tl.float32)
+    # compute
+    delta = tl.sum(o * do, axis=1)
+    # write-back, shape (q.shape[0] * q.shape[1], q.shape[2])
+    off_zh = off_z * num_h + off_h * 1
+    # Check for OOB accesses
+    delta_ptrs = Delta + off_zh * seqlen_q + off_m + tl.arange(0, BLOCK_M)
+    overflow = off_m + BLOCK_M - seqlen_q
+    if overflow > 0:
+        boundary = tl.full((BLOCK_M, ), BLOCK_M - overflow, dtype=tl.int32)
+        mask = boundary > tl.arange(0, BLOCK_M)
+        tl.store(delta_ptrs, delta, mask=mask)
+    else:
+        tl.store(delta_ptrs, delta)
+
+@triton.jit
+def _bwd_kernel_dk_dv(
+                   dk, dv,
+                   Q, k, v, sm_scale, alibi_slope,
+                   DO,
+                   M, D,
+                   # shared by Q/K/V/DO.
+                   stride_tok, stride_d,
+                   H, N_CTX, BLOCK_M1: tl.constexpr,
+                   BLOCK_N1: tl.constexpr,
+                   BLOCK_DMODEL: tl.constexpr,
+                   # Filled in by the wrapper.
+                   start_n, start_m, num_steps,
+                   MASK: tl.constexpr):
+    offs_m = start_m + tl.arange(0, BLOCK_M1)
+    offs_n = start_n + tl.arange(0, BLOCK_N1)
+    offs_k = tl.arange(0, BLOCK_DMODEL)
+    QT_block_ptr = tl.make_block_ptr(
+        base=Q,
+        shape=(BLOCK_DMODEL, N_CTX),
+        strides=(stride_d, stride_tok),
+        offsets=(0, start_m),
+        block_shape=(BLOCK_DMODEL, BLOCK_M1),
+        order=(0,1)
+    )
+    DO_block_ptr = tl.make_block_ptr(
+        base=DO,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_m, 0),
+        block_shape=(BLOCK_M1, BLOCK_DMODEL),
+        order=(1,0)
+    )
+    # BLOCK_N1 must be a multiple of BLOCK_M1, otherwise the code wouldn't work.
+    tl.static_assert(BLOCK_N1 % BLOCK_M1 == 0)
+    curr_m = start_m
+    step_m = BLOCK_M1
+    for blk_idx in range(num_steps):
+        qT = tl.load(QT_block_ptr)
+        # Load m before computing qk to reduce pipeline stall.
+        offs_m = curr_m + tl.arange(0, BLOCK_M1)
+        m = tl.load(M + offs_m)
+        kqT = tl.dot(k, qT)
+        if alibi_slope is not None:
+            alibi_block = compute_alibi_block(alibi_slope, N_CTX, N_CTX, offs_m, offs_n, True)
+            kqT += alibi_block * 1.44269504089
+
+        pT = tl.math.exp2(kqT - m[None, :])
+        # Autoregressive masking.
+        if MASK:
+            mask = (offs_m[None, :] >= offs_n[:, None])
+            pT = tl.where(mask, pT, 0.0)
+        do = tl.load(DO_block_ptr)
+        # Compute dV.
+        ppT = pT
+        ppT = ppT.to(tl.float16)
+        dv += tl.dot(ppT, do)
+        # D (= delta) is pre-divided by ds_scale.
+        Di = tl.load(D + offs_m)
+        # Compute dP and dS.
+        dpT = tl.dot(v, tl.trans(do))
+        dsT = pT * (dpT - Di[None, :])
+        dsT = dsT.to(tl.float16)
+        dk += tl.dot(dsT, tl.trans(qT))
+        # Increment pointers.
+        curr_m += step_m
+        QT_block_ptr = tl.advance(QT_block_ptr, (0, step_m))
+        DO_block_ptr = tl.advance(DO_block_ptr, (step_m, 0))
+    return dk, dv
+
+@triton.jit
+def _bwd_kernel_dq(dq, q, K, V,
+                 do, m, D, alibi_slope,
+                 # shared by Q/K/V/DO.
+                 stride_tok, stride_d,
+                 H, N_CTX,
+                 BLOCK_M2: tl.constexpr,
+                 BLOCK_N2: tl.constexpr,
+                 BLOCK_DMODEL: tl.constexpr,
+                 # Filled in by the wrapper.
+                 start_m, start_n, num_steps,
+                 MASK: tl.constexpr):
+    offs_m = start_m + tl.arange(0, BLOCK_M2)
+    offs_n = start_n + tl.arange(0, BLOCK_N2)
+    offs_k = tl.arange(0, BLOCK_DMODEL)
+    KT_block_ptr = tl.make_block_ptr(
+        base=K,
+        shape=(BLOCK_DMODEL, N_CTX),
+        strides=(stride_d, stride_tok),
+        offsets=(0, start_n),
+        block_shape=(BLOCK_DMODEL, BLOCK_N2),
+        order=(0, 1)
+    )
+    VT_block_ptr = tl.make_block_ptr(
+        base=V,
+        shape=(BLOCK_DMODEL, N_CTX),
+        strides=(stride_d, stride_tok),
+        offsets=(0, start_n),
+        block_shape=(BLOCK_DMODEL, BLOCK_N2),
+        order=(0, 1)
+    )
+    # D (= delta) is pre-divided by ds_scale.
+    Di = tl.load(D + offs_m)
+    # BLOCK_M2 must be a multiple of BLOCK_N2, otherwise the code wouldn't work.
+    tl.static_assert(BLOCK_M2 % BLOCK_N2 == 0)
+    curr_n = start_n
+    step_n = BLOCK_N2
+    for blk_idx in range(num_steps):
+        kT = tl.load(KT_block_ptr)
+        qk = tl.dot(q, kT)
+        if alibi_slope is not None:  
+            alibi_block = compute_alibi_block(alibi_slope, N_CTX, N_CTX, offs_m, offs_n)
+            qk += alibi_block * 1.44269504089
+
+        p = tl.math.exp2(qk - m)
+        # Autoregressive masking.
+        if MASK:
+            offs_n = curr_n + tl.arange(0, BLOCK_N2)
+            mask = (offs_m[:, None] >= offs_n[None, :])
+            p = tl.where(mask, p, 0.0)
+        # Compute dP and dS.
+        vT = tl.load(VT_block_ptr)
+        dp = tl.dot(do, vT).to(tl.float32)
+        ds = p * (dp - Di[:, None])
+        ds = ds.to(tl.float16)
+        # Compute dQ.0.
+        # NOTE: We need to de-scale dq in the end, because kT was pre-scaled.
+        dq += tl.dot(ds, tl.trans(kT))
+        # Increment pointers.
+        curr_n += step_n
+        KT_block_ptr = tl.advance(KT_block_ptr, (0, step_n))
+        VT_block_ptr = tl.advance(VT_block_ptr, (0, step_n))
+    return dq
+
+@triton.jit
+def _attn_bwd(Q, K, V, sm_scale, alibi_slopes,
+              DO,
+              DQ, DK, DV,
+              M, D,
+              # shared by Q/K/V/DO.
+              stride_z, stride_h, stride_tok, stride_d,
+              # H = 16, N_CTX = 1024
+              H, N_CTX,
+              BLOCK_DMODEL: tl.constexpr,
+              BLOCK_M1: tl.constexpr,
+              BLOCK_N1: tl.constexpr,
+              BLOCK_M2: tl.constexpr,
+              BLOCK_N2: tl.constexpr,
+              BLK_SLICE_FACTOR: tl.constexpr,
+              USE_ALIBI: tl.constexpr):
+    LN2: tl.constexpr = 0.6931471824645996  # = ln(2)
+
+    bhid = tl.program_id(2)
+    off_chz = (bhid * N_CTX).to(tl.int64)
+    adj = (stride_h * (bhid % H) + stride_z * (bhid // H)).to(tl.int64)
+    pid = tl.program_id(0)
+
+    # offset pointers for batch/head
+    Q += adj
+    K += adj
+    V += adj
+    DO += adj
+    DQ += adj
+    DK += adj
+    DV += adj
+    M += off_chz
+    D += off_chz
+
+    offs_k = tl.arange(0, BLOCK_DMODEL)
+
+    start_n = pid * BLOCK_N1
+    # This assignment is important. It is what allows us to pick the diagonal
+    # blocks. Later, when we want to do the lower triangular, we update start_m
+    # after the first dkdv call.
+    start_m = start_n
+
+    MASK_BLOCK_M1: tl.constexpr = BLOCK_M1 // BLK_SLICE_FACTOR
+    offs_n = start_n + tl.arange(0, BLOCK_N1)
+
+    dv = tl.zeros([BLOCK_N1, BLOCK_DMODEL], dtype=tl.float32)
+    dk = tl.zeros([BLOCK_N1, BLOCK_DMODEL], dtype=tl.float32)
+
+    K_block_ptr = tl.make_block_ptr(
+        base=K,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_n, 0),
+        block_shape=(BLOCK_N1, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    V_block_ptr = tl.make_block_ptr(
+        base=V,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_n, 0),
+        block_shape=(BLOCK_N1, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+
+    # load K and V: they stay in SRAM throughout the inner loop for dkdv.
+    k = tl.load(K_block_ptr)
+    v = tl.load(V_block_ptr)
+
+    if USE_ALIBI:
+        a_offset = bhid
+        alibi_slope = tl.load(alibi_slopes + a_offset)
+    else:
+        alibi_slope = None
+
+    # compute dK and dV for blocks close to the diagonal that need to be masked
+    num_steps = BLOCK_N1 // MASK_BLOCK_M1
+    dk, dv = _bwd_kernel_dk_dv(
+                            dk, dv,
+                            Q, k, v, sm_scale, alibi_slope,
+                            DO,
+                            M, D,
+                            stride_tok, stride_d,
+                            H, N_CTX,
+                            MASK_BLOCK_M1, BLOCK_N1, BLOCK_DMODEL,
+                            start_n, start_m, num_steps,
+                            MASK=True
+                            )
+
+    # compute dK and dV for blocks that don't need masking further from the diagonal
+    start_m += num_steps * MASK_BLOCK_M1
+    num_steps = (N_CTX - start_m) // BLOCK_M1
+
+    dk, dv = _bwd_kernel_dk_dv(
+        dk, dv,
+        Q, k, v, sm_scale, alibi_slope,
+        DO,
+        M, D,
+        stride_tok, stride_d,
+        H, N_CTX,
+        BLOCK_M1, BLOCK_N1, BLOCK_DMODEL,
+        start_n, start_m, num_steps,
+        MASK=False
+    )
+
+    DV_block_ptrs = tl.make_block_ptr(
+        base=DV,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_n, 0),
+        block_shape=(BLOCK_N1, BLOCK_DMODEL),
+        order=(1,0)
+    )
+    tl.store(DV_block_ptrs, dv.to(v.dtype))
+
+    # Write back dK.
+    dk *= sm_scale
+    DK_block_ptrs = tl.make_block_ptr(
+        base=DK,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_n, 0),
+        block_shape=(BLOCK_N1, BLOCK_DMODEL),
+        order=(1,0)
+    )
+    tl.store(DK_block_ptrs, dk.to(k.dtype))
+
+    # THIS BLOCK DOES DQ:
+    start_m = pid * BLOCK_M2
+    end_n = start_m + BLOCK_M2
+
+    MASK_BLOCK_N2: tl.constexpr = BLOCK_N2 // BLK_SLICE_FACTOR
+    offs_m = start_m + tl.arange(0, BLOCK_M2)
+
+    Q_block_ptr = tl.make_block_ptr(
+        base=Q,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_m, 0),
+        block_shape=(BLOCK_M2, BLOCK_DMODEL),
+        order=(1, 0)
+    )
+
+    DO_block_ptr = tl.make_block_ptr(
+        base=DO,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_m, 0),
+        block_shape=(BLOCK_M2, BLOCK_DMODEL),
+        order=(1, 0)
+    )
+    q = tl.load(Q_block_ptr)
+    do = tl.load(DO_block_ptr)
+    dq = tl.zeros([BLOCK_M2, BLOCK_DMODEL], dtype=tl.float32)
+
+    m = tl.load(M + offs_m)
+    m = m[:, None]
+
+    # Compute dQ for masked (diagonal) blocks.
+    # NOTE: This code scans each row of QK^T backward (from right to left,
+    # but inside each call to _attn_bwd_dq, from left to right), but that's
+    # not due to anything important.  I just wanted to reuse the loop
+    # structure for dK & dV above as much as possible.
+    num_steps = BLOCK_M2 // MASK_BLOCK_N2
+    dq = _bwd_kernel_dq(dq, q, K, V,
+                      do, m, D, alibi_slope,
+                      stride_tok, stride_d,
+                      H, N_CTX,
+                      BLOCK_M2, MASK_BLOCK_N2, BLOCK_DMODEL,
+                      start_m, end_n - num_steps * MASK_BLOCK_N2, num_steps,
+                      MASK=True
+                      )
+    end_n -= num_steps * MASK_BLOCK_N2
+    # stage 2
+    num_steps = end_n // BLOCK_N2
+    dq = _bwd_kernel_dq(dq, q, K, V,
+                      do, m, D, alibi_slope,
+                      stride_tok, stride_d,
+                      H, N_CTX,
+                      BLOCK_M2, BLOCK_N2, BLOCK_DMODEL,
+                      start_m, end_n - num_steps * BLOCK_N2, num_steps,
+                      MASK=False
+                      )
+    # Write back dQ.
+    DQ_block_ptr = tl.make_block_ptr(
+        base=DQ,
+        shape=(N_CTX, BLOCK_DMODEL),
+        strides=(stride_tok, stride_d),
+        offsets=(start_m, 0),
+        block_shape=(BLOCK_M2, BLOCK_DMODEL),
+        order=(1, 0)
+    )
+    dq *= LN2
+    tl.store(DQ_block_ptr, dq.to(q.dtype))
+
+class _attention(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, q, k, v, o, metadata):
+        # NOTE: a large bias tensor leads to overflow during pointer arithmetic
+        if (metadata.bias is not None):
+            assert(metadata.bias.numel() < 2 ** 31)    
+
+        if o is None:
+            o = torch.empty_like(q, dtype=v.dtype)
+        import os
+        if os.environ.get("FLASH_ATTENTION_PRINT_PARAM", "0") == "1":
+            print(f"triton flash attention: {q.shape=}, {k.shape=}, {v.shape}, {o.shape=}")
+            print(f"triton flash attention: {q.stride()=}, {k.stride()=}, {v.stride()=}, {o.stride()=}")
+            print(f"triton flash attention: {metadata=}")
+        metadata.check_args(q, k, v, o)
+        if metadata.varlen:
+            total_q, nheads_q, head_size = q.shape
+            total_k, nheads_k, _ = k.shape
+            batch = metadata.num_contexts
+            q_strides = (0, q.stride(1), q.stride(0), q.stride(2))
+            k_strides = (0, k.stride(1), k.stride(0), k.stride(2))
+            v_strides = (0, v.stride(1), v.stride(0), v.stride(2))
+            o_strides = (0, o.stride(1), o.stride(0), o.stride(2))
+        else:
+            batch, nheads_q, seqlen_q, head_size = q.shape
+            _, nheads_k, seqlen_k, _ = k.shape
+            q_strides = (q.stride(0), q.stride(1), q.stride(2), q.stride(3))
+            k_strides = (k.stride(0), k.stride(1), k.stride(2), k.stride(3))
+            v_strides = (v.stride(0), v.stride(1), v.stride(2), v.stride(3))
+            o_strides = (o.stride(0), o.stride(1), o.stride(2), o.stride(3))
+
+        # Get closest power of 2 over or equal to 32.
+        padded_d_model = 1 << (head_size - 1).bit_length()
+        padded_d_model = max(padded_d_model, 16)
+
+        grid = lambda META: (
+            triton.cdiv(metadata.max_seqlens_q, META['BLOCK_M']),
+            nheads_q,
+            batch
+        )
+
+        # encoded_softmax is used to validate dropout behavior vs the PyTorch SDPA math backend reference.  We zero this out
+        # to give a consistent starting point and then populate it with the output of softmax with the sign bit set according
+        # to the dropout mask. The resulting return allows this mask to be fed into the reference implementation for testing
+        # only.  This return holds no useful output aside from debugging.
+        if metadata.return_encoded_softmax:
+            encoded_softmax = torch.zeros((q.shape[0], q.shape[1], q.shape[2], k.shape[2]), device=q.device, dtype=torch.float32)
+        else:
+            encoded_softmax = None
+
+        M = torch.empty((batch, nheads_q, metadata.max_seqlens_q), device=q.device, dtype=torch.float32)
+
+        # Seed the RNG so we get reproducible results for testing.
+        philox_seed = 0x1BF52
+        philox_offset = 0x1D4B42
+
+        if metadata.bias is not None:
+            bias_strides = (metadata.bias.stride(0), metadata.bias.stride(1),
+                            metadata.bias.stride(2), metadata.bias.stride(3))
+        else:
+            bias_strides = (0,0,0,0)
+        
+        if metadata.alibi_slopes is not None:
+            alibi_strides = (metadata.alibi_slopes.stride(0), metadata.alibi_slopes.stride(1))
+        else:
+            alibi_strides = (0, 0)
+
+        attn_fwd[grid](
+            q, k, v, metadata.bias, metadata.sm_scale, M, o,
+            *q_strides, *k_strides, *v_strides, *o_strides, *bias_strides, *alibi_strides,
+            metadata.cu_seqlens_q, metadata.cu_seqlens_k,
+            dropout_p=metadata.dropout_p,
+            philox_seed=philox_seed,
+            philox_offset_base=philox_offset,
+            encoded_softmax=encoded_softmax,
+            alibi_slopes = metadata.alibi_slopes,
+            HQ=nheads_q, HK=nheads_k,
+            ACTUAL_BLOCK_DMODEL=head_size,
+            MAX_SEQLENS_Q=metadata.max_seqlens_q,
+            MAX_SEQLENS_K=metadata.max_seqlens_k,
+            IS_CAUSAL=metadata.causal,
+            VARLEN=metadata.varlen,
+            BLOCK_DMODEL=padded_d_model,
+            BIAS_TYPE=0 if metadata.bias is None else 1,
+            USE_ALIBI=False if metadata.alibi_slopes is None else True,
+            ENABLE_DROPOUT=metadata.dropout_p > 0.0,
+            RETURN_ENCODED_SOFTMAX=metadata.return_encoded_softmax,
+            BATCH_SIZE= q.shape[0]
+        )
+        if os.environ.get("FLASH_ATTENTION_PRINT_PARAM", "0") == "1":
+            best_config = attn_fwd.get_best_config()
+            print(f"{best_config.kwargs=}, {best_config.num_stages=}, {best_config.num_warps=}")
+
+        ctx.save_for_backward(q, k, v, o, M)
+        ctx.grid = grid
+        ctx.sm_scale = metadata.sm_scale
+        ctx.BLOCK_DMODEL = head_size
+        ctx.causal = metadata.causal
+        ctx.alibi_slopes = metadata.alibi_slopes
+        ctx.dropout_p = metadata.dropout_p
+        ctx.philox_seed = philox_seed
+        ctx.philox_offset = philox_offset
+        ctx.encoded_softmax = encoded_softmax
+        ctx.return_encoded_softmax = metadata.return_encoded_softmax
+        return o if not metadata.return_encoded_softmax else (o, encoded_softmax, ) # S_dmask
+
+    @staticmethod
+    def backward(ctx, do, *args):
+        if torch.version.hip is not None:
+            BLOCK = 64
+        else:
+            BLOCK = 128
+        q, k, v, o, M = ctx.saved_tensors
+        import os
+        if os.environ.get("TRITON_FLASHATTN_DEBUG", "0") == "1":
+            print(f"triton flash attention: {q.shape=}, {k.shape=}, {v.shape}, {o.shape=}, {do.shape=}")
+            print(f"triton flash attention: {q.stride()=}, {k.stride()=}, {v.stride()=}, {o.stride()=}, {do.stride()}")
+        # assert do.is_contiguous()
+        assert q.stride() == k.stride() == v.stride() == o.stride() == do.stride()
+        seqlen_q = q.shape[2]
+        dq = torch.empty_like(q)
+        dk = torch.empty_like(k)
+        dv = torch.empty_like(v)
+        BATCH, N_HEAD, N_CTX = q.shape[:3]
+        PRE_BLOCK = 128
+        NUM_WARPS, NUM_STAGES = 4, 1
+        BLOCK_M1, BLOCK_N1, BLOCK_M2, BLOCK_N2 = 32, 64, 64, 32
+        BLK_SLICE_FACTOR = 2
+        RCP_LN2 = 1.4426950408889634  # = 1.0 / ln(2)
+        arg_k = k
+        arg_k = arg_k * (ctx.sm_scale * RCP_LN2)
+        assert N_CTX % PRE_BLOCK == 0
+        delta = torch.empty_like(M)
+        Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
+        padded_head = (Lk != ctx.BLOCK_DMODEL)
+        grid_preprocess = (triton.cdiv(do.shape[2], BLOCK), do.shape[1], do.shape[0])
+        _attn_bwd_preprocess[grid_preprocess](
+            o, do, delta,
+            o.stride(0), o.stride(1), o.stride(2), o.stride(3),
+            do.stride(0), do.stride(1), do.stride(2), do.stride(3),
+            seqlen_q,
+            head_dim=Lk,
+            BLOCK_M=BLOCK, D_HEAD=ctx.BLOCK_DMODEL,
+        )
+        grid = lambda META: (
+            triton.cdiv(N_CTX, META['BLOCK_N1']),
+            1,
+            BATCH * N_HEAD
+        )
+        _attn_bwd[grid](
+            q, arg_k, v, ctx.sm_scale, ctx.alibi_slopes, do, dq, dk, dv,
+            M, delta,
+            q.stride(0), q.stride(1), q.stride(2), q.stride(3),
+            N_HEAD, N_CTX,
+            BLOCK_DMODEL=ctx.BLOCK_DMODEL,
+            BLOCK_M1=BLOCK_M1, BLOCK_N1=BLOCK_N1, BLOCK_M2=BLOCK_M2, BLOCK_N2=BLOCK_N2,
+            BLK_SLICE_FACTOR=BLK_SLICE_FACTOR,
+            USE_ALIBI= False if ctx.alibi_slopes is None else True,
+        )
+
+        return dq, dk, dv, None, None
+
+attention = _attention.apply
+
+# flash_attn wrapper
+
+def input_helper(Z, HQ, HK, N_CTX_Q, N_CTX_K, D_HEAD, dtype):
+    torch.manual_seed(20)
+
+    # Initialize q, k, v
+    q = torch.randn((Z, HQ, N_CTX_Q, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
+    k = torch.randn((Z, HK, N_CTX_K, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
+    v = torch.randn((Z, HK, N_CTX_K, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
+    sm_scale = D_HEAD**-0.5
+    input_metadata = MetaData(sm_scale=sm_scale)
+    input_metadata.max_seqlens_q = N_CTX_Q
+    input_metadata.max_seqlens_k = N_CTX_K
+    return q, k, v, input_metadata
+
+def padding_bshd(t):   # BSHD 
+    batch, seqlen, nheads, dim = t.shape
+    t = torch.nn.functional.pad(t.reshape(batch, seqlen, nheads*dim), (0, 32), 'constant', 0)[:,:,:-32].reshape(batch, seqlen, nheads, dim) # pad: nheads*dim+32
+    # t = torch.nn.functional.pad(t.reshape(batch, seqlen, nheads, dim), (0, 32), 'constant', 0)[:,:,:,:-32] # pad: dim+32
+    return t
+
+def flash_attn_func(q, k, v, dropout_p=0.0, softmax_scale=None, causal=False,
+                    return_attn_probs=False, padding_input=False):
+    if padding_input:
+        k, v = (padding_bshd(t) for t in (k, v))
+    q, k, v = (t.transpose(1, 2) for t in (q, k, v))
+    softmax_scale = softmax_scale if softmax_scale else q.shape[-1]**-0.5
+    input_metadata = MetaData(sm_scale=softmax_scale, causal=causal, dropout_p=dropout_p, return_encoded_softmax=return_attn_probs)
+    input_metadata.max_seqlens_q = q.shape[2]
+    input_metadata.max_seqlens_k = k.shape[2]
+    return _attention.apply(q, k, v, None, input_metadata).transpose(1, 2)
+
+def flash_attn_kvpacked_func(q, kv, dropout_p=0.0, softmax_scale=None, causal=False,
+                    return_attn_probs=False, padding_input=False):
+    k, v = kv[:, :, 0], kv[:, :, 1]  # batch_size, seqlen, 2, nheads_k, headdim
+    if padding_input:
+        k, v = (padding_bshd(t) for t in (k, v))   #  pad 
+    q, k, v = (t.transpose(1, 2) for t in (q, k, v)) # trans bshd to bhsd
+    softmax_scale = softmax_scale if softmax_scale else q.shape[-1]**-0.5
+    input_metadata = MetaData(sm_scale=softmax_scale, causal=causal, dropout_p=dropout_p, return_encoded_softmax=return_attn_probs)
+    input_metadata.max_seqlens_q = q.shape[2]
+    input_metadata.max_seqlens_k = k.shape[2]
+    return _attention.apply(q, k, v, None, input_metadata).transpose(1, 2) # trans bhsd to bshd
+
+def flash_attn_qkvpacked_func(qkv, dropout_p=0.0, softmax_scale=None, causal=False,
+                              return_attn_probs=False, padding_input=False):
+    q, k, v = qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2]
+    if padding_input:
+        k, v = (padding_bshd(t) for t in (k, v))
+    q, k, v = (t.transpose(1, 2) for t in (q, k, v))
+    softmax_scale = softmax_scale if softmax_scale else q.shape[-1]**-0.5
+    input_metadata = MetaData(sm_scale=softmax_scale, causal=causal, dropout_p=dropout_p, return_encoded_softmax=return_attn_probs)
+    input_metadata.max_seqlens_q = q.shape[2]
+    input_metadata.max_seqlens_k = k.shape[2]
+    return _attention.apply(q, k, v, None, input_metadata).transpose(1, 2)
+
+# varlen flash_attn
+
+def varlen_input_helper(Z, HQ, HK, N_CTX_Q, N_CTX_K, D_HEAD, dtype, causal):
+    torch.manual_seed(20)
+
+    # Random sequence lengths. Using N_CTX as kind of max of sum of individual seqs
+    max_seqlens_q = N_CTX_Q // Z
+    max_seqlens_k = N_CTX_K // Z
+    seqlens_q = torch.randint(1, max_seqlens_q + 1, (Z,), dtype=torch.int32)
+    seqlens_k = torch.randint(1, max_seqlens_k + 1, (Z,), dtype=torch.int32)
+    max_seqlens_q = torch.max(seqlens_q).item()
+    max_seqlens_k = torch.max(seqlens_k).item()
+
+    # Calculate cumulative sequence lengths
+    cu_seqlens_q = torch.cat([torch.tensor([0], dtype=torch.int32), seqlens_q.cumsum(dim=0, dtype=torch.int32)])
+    cu_seqlens_k = torch.cat([torch.tensor([0], dtype=torch.int32), seqlens_k.cumsum(dim=0, dtype=torch.int32)])
+    cu_seqlens_q = cu_seqlens_q.to(device="cuda")
+    cu_seqlens_k = cu_seqlens_k.to(device="cuda")
+    # -1 because the last entry of cu_seqlens_q specifies the end of the last seq
+    num_ctxs = len(cu_seqlens_q) - 1
+
+    # Initialize q, k, v with variable lengths
+    total_q = cu_seqlens_q[-1].item()
+    total_k = cu_seqlens_k[-1].item()
+    q = torch.randn((total_q, HQ, D_HEAD), dtype=dtype, device="cuda").normal_(mean=0., std=0.5).requires_grad_()
+    k = torch.randn((total_k, HK, D_HEAD), dtype=dtype, device="cuda").normal_(mean=0., std=0.5).requires_grad_()
+    v = torch.randn((total_k, HK, D_HEAD), dtype=dtype, device="cuda").normal_(mean=0., std=0.5).requires_grad_()
+    sm_scale = D_HEAD**-0.5
+    input_metadata = MetaData(sm_scale=sm_scale)
+    input_metadata.set_varlen_params(cu_seqlens_q, cu_seqlens_k)
+    input_metadata.max_seqlens_q = max_seqlens_q
+    input_metadata.max_seqlens_k = max_seqlens_k
+    if causal:
+        input_metadata.need_causal()
+    return q, k, v, input_metadata
+
+def padding_thd(t):   # THD 
+    total_seqlen, nheads, dim = t.shape
+    t = torch.nn.functional.pad(t.reshape(total_seqlen, nheads*dim), (0, 32), 'constant', 0)[:,:-32].reshape(total_seqlen, nheads, dim) # pad: nheads*dim+32
+    # t = torch.nn.functional.pad(t.reshape(total_seqlen, nheads, dim), (0, 32), 'constant', 0)[:,:-32] # pad: dim+32
+    return t
+
+def flash_attn_varlen_qkvpacked_func(qkv, cu_seqlens, max_seqlens, dropout_p=0.0, softmax_scale=None,
+                                     causal=False, return_attn_probs=False, padding_input=False):
+    q, k, v = qkv[:, 0], qkv[:, 1], qkv[:, 2]  # total_seqlen, 3, nheads, dim
+    if padding_input:
+        k, v = (padding_thd(t) for t in (k, v))   #  pad 
+    softmax_scale = softmax_scale if softmax_scale else q.shape[-1]**-0.5
+    input_metadata = MetaData(sm_scale=softmax_scale, causal=causal, dropout_p=dropout_p, return_encoded_softmax=return_attn_probs)
+    input_metadata.set_varlen_params(cu_seqlens, cu_seqlens)
+    input_metadata.max_seqlens_q = max_seqlens
+    input_metadata.max_seqlens_k = max_seqlens
+    return  _attention.apply(q, k, v, None, input_metadata)
+
+
+def flash_attn_varlen_kvpacked_func(q, kv, cu_seqlens_q, cu_seqlens_k, max_seqlens_q, max_seqlens_k,
+                                    dropout_p=0.0, softmax_scale=None, causal=False,
+                                    return_attn_probs=False, padding_input=False):
+    k, v = kv[:, 0], kv[:, 1]  # total_seqlen, 2, nheads, dim
+    if padding_input:
+        k, v = (padding_thd(t) for t in (k, v))
+    softmax_scale = softmax_scale if softmax_scale else q.shape[-1]**-0.5
+    input_metadata = MetaData(sm_scale=softmax_scale, causal=causal, dropout_p=dropout_p, return_encoded_softmax=return_attn_probs)
+    input_metadata.set_varlen_params(cu_seqlens_q, cu_seqlens_k)
+    input_metadata.max_seqlens_q = max_seqlens_q
+    input_metadata.max_seqlens_k = max_seqlens_k
+    return  _attention.apply(q, k, v, None, input_metadata)
+
+
+def flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlens_q, max_seqlens_k,
+                           dropout_p=0.0, softmax_scale=None, causal=False,
+                           return_attn_probs=False, padding_input=False):
+    if padding_input:
+        k, v = (padding_thd(t) for t in (k, v))
+    softmax_scale = softmax_scale if softmax_scale else q.shape[-1]**-0.5
+    input_metadata = MetaData(sm_scale=softmax_scale, causal=causal, dropout_p=dropout_p, return_encoded_softmax=return_attn_probs)
+    input_metadata.set_varlen_params(cu_seqlens_q, cu_seqlens_k)
+    input_metadata.max_seqlens_q = max_seqlens_q
+    input_metadata.max_seqlens_k = max_seqlens_k
+    return  _attention.apply(q, k, v, None, input_metadata)
+
+# legacy interface
+def flash_attn_unpadded_qkvpacked_func(qkv, cu_seqlens, max_seqlens, dropout_p, softmax_scale=None,
+                                       causal=False, return_attn_probs=False, padding_input=False):
+  return flash_attn_varlen_qkvpacked_func(qkv, cu_seqlens, max_seqlens, dropout_p, softmax_scale,
+                                          causal, return_attn_probs, padding_input)
+  
+def flash_attn_unpadded_kvpacked_func(q, kv, cu_seqlens_q, cu_seqlens_k, max_seqlens_q, max_seqlens_k,
+                                      dropout_p, softmax_scale=None, causal=False, return_attn_probs=False, padding_input=False):
+  return flash_attn_varlen_kvpacked_func(q, kv, cu_seqlens_q, cu_seqlens_k, max_seqlens_q, max_seqlens_k,
+                                         dropout_p, softmax_scale, causal, return_attn_probs, padding_input)
+    
+def flash_attn_unpadded_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlens_q, max_seqlens_k,
+                             dropout_p, softmax_scale=None, 
+                             causal=False, return_attn_probs=False, padding_input=False):
+  return flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlens_q, max_seqlens_k,
+                                dropout_p, softmax_scale, causal, return_attn_probs, padding_input)
+
+

vllm/envs.py

@@ -130,7 +130,7 @@ environment_variables: Dict[str, Callable[[], Any]] = {
 
     # flag to control if vllm should use triton flash attention
     "VLLM_USE_TRITON_FLASH_ATTN":
-    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "False").lower() in
+    lambda: (os.environ.get("VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in
              ("true", "1")),
 
     # local rank of the process in the distributed setting, used to determine