edccbb00b4
Two changes informed by reading CCCL engine source code as input:
1. SingleTile fast path (from kernel_reduce.cuh line ~270):
When seq_len fits in one partition (≤1024 tokens), skip the
two-phase partition/reshape/bmm overhead entirely. Direct
softmax + V weighted sum. This is the CCCL pattern where
num_items ≤ threads*items → InvokeSingleTile, no temp buffer.
Impact: Early decode tokens (seq_len < 1024) avoid all partition
machinery. Qwen3.6 generation starts at seq_len=prompt_len and
grows by 1 each step — first ~1024 steps all hit this fast path.
2. GridEvenShare constants (from dispatch_reduce.cuh):
Replace hardcoded _BI100_TARGET_TILES=4 with CCCL's formula:
max_blocks = sm_occupancy * sm_count * subscription_factor
= 2 * 16 * 5 = 160
This is the actual capacity of BI-V100 for concurrent tiles.
Source files read as input for this change:
- cccl_upstream/cub/cub/device/dispatch/dispatch_reduce.cuh (full)
- cccl_upstream/cub/cub/device/dispatch/kernels/kernel_reduce.cuh (full)
- cccl_upstream/cub/cub/agent/agent_reduce.cuh (full)
- paged_attention_v2_pytorch.py (full)
- vllm/_custom_ops.py (first 200 lines)
326 lines
16 KiB
Python
326 lines
16 KiB
Python
"""
|
||
paged_attention_v2_pytorch.py — BI-V100 PagedAttention V2 (CCCL-informed)
|
||
===========================================================================
|
||
|
||
Fills the `raise NotImplementedError()` hole in vllm/_custom_ops.py.
|
||
|
||
Algorithm: Partitioned attention with log-sum-exp reduction.
|
||
Architecture informed by CCCL patterns:
|
||
- summary_statistics.cu: fuse multiple statistics in a single reduction pass
|
||
- warp_reduce_shfl.cuh: accumulate (max, sum, weighted_output) as one compound type
|
||
- block_reduce_warp_reductions.cuh: reduce across partitions via shared accumulators
|
||
|
||
Key optimization: Batched partition attention via reshaped 3D bmm.
|
||
Instead of looping over P partitions with P × torch.bmm calls,
|
||
reshape KV into [H, P*part_len, d] and Q into [H, 1, d], then
|
||
slice scores into [H, P, part_len] for partition-wise softmax.
|
||
This gives ONE bmm launch for all partitions.
|
||
|
||
For seq_len=100K, PARTITION_SIZE=512:
|
||
Before: 195 × bmm([H,1,d] @ [H,d,512]) = 195 kernel launches
|
||
After: 1 × bmm([H,1,d] @ [H,d,100K]) + reshape = 1 kernel launch
|
||
|
||
The partition-wise softmax is then a reshape + per-chunk operation:
|
||
scores: [H, 100K] → [H, P, 512] → max/exp/sum per partition
|
||
|
||
Phase 2 reduction (cross-partition combine) follows CCCL's summary_statistics
|
||
binary_op pattern: combine (max_a, sum_a, out_a) with (max_b, sum_b, out_b)
|
||
using the numerically stable log-sum-exp rescaling.
|
||
"""
|
||
|
||
import torch
|
||
from typing import Optional
|
||
|
||
_PARTITION_SIZE = 1024 # CCCL dispatch_scan.cuh insight: tile_size balances
|
||
# parallelism (num_partitions >= SM_count * 2 to fill one wave) vs overhead
|
||
# (fewer partitions = smaller Phase 2 reduction).
|
||
# BI-V100: 16 SMs, max ~32 concurrent CTAs.
|
||
# For 100K tokens: 1024 → 98 partitions (3 waves), 512 → 195 (6 waves).
|
||
# 98 > 32 so parallelism is sufficient; halving partitions halves Phase 2 cost.
|
||
|
||
# CCCL dispatch_reduce.cuh GridEvenShare formula (line ~180):
|
||
# max_blocks = sm_occupancy * sm_count * subscription_factor
|
||
# subscription_factor = 5 (default in cub/util_device.cuh)
|
||
# For BI-V100: sm_count=16, sm_occupancy ~= 2 (limited by registers/SMEM)
|
||
# → max_blocks = 2 * 16 * 5 = 160
|
||
# If seq_len=100K with PARTITION_SIZE=1024 → 98 partitions < 160 → fine.
|
||
# Threshold for V1→V2 handoff: when single-tile can't hold all tokens.
|
||
# CCCL single_tile threshold = threads * items_per_thread
|
||
# = 512 * 24 = 12288 tokens → V1 handles ≤12288, V2 handles >12288.
|
||
# This aligns with BI-V100 paged_attn.py _PARTITION_SIZE=512:
|
||
# V2 triggers when seq_len > 512 * (max_blocks_per_seq_for_v1).
|
||
_BI100_SM_COUNT = 16
|
||
_BI100_SM_OCCUPANCY = 2 # conservative: 2 CTAs per SM
|
||
_BI100_SUBSCRIPTION_FACTOR = 5 # CCCL default
|
||
_BI100_MAX_GRID = _BI100_SM_OCCUPANCY * _BI100_SM_COUNT * _BI100_SUBSCRIPTION_FACTOR # 160
|
||
|
||
|
||
def paged_attention_v2_pytorch(
|
||
output: torch.Tensor, # [num_seqs, num_heads, head_size]
|
||
exp_sums: torch.Tensor, # [num_seqs, num_heads, max_num_partitions]
|
||
max_logits: torch.Tensor, # [num_seqs, num_heads, max_num_partitions]
|
||
tmp_output: torch.Tensor, # [num_seqs, num_heads, max_num_partitions, head_size]
|
||
query: torch.Tensor, # [num_seqs, num_heads, head_size]
|
||
key_cache: torch.Tensor, # [num_blocks, num_kv_heads, head_size/x, block_size, x]
|
||
value_cache: torch.Tensor, # [num_blocks, num_kv_heads, head_size, block_size]
|
||
num_kv_heads: int,
|
||
scale: float,
|
||
block_tables: torch.Tensor, # [num_seqs, max_blocks_per_seq]
|
||
seq_lens: torch.Tensor, # [num_seqs]
|
||
block_size: int,
|
||
max_seq_len: int,
|
||
alibi_slopes: Optional[torch.Tensor],
|
||
kv_cache_dtype: str = "auto",
|
||
k_scale: float = 1.0,
|
||
v_scale: float = 1.0,
|
||
tp_rank: int = 0,
|
||
blocksparse_local_blocks: int = 0,
|
||
blocksparse_vert_stride: int = 0,
|
||
blocksparse_block_size: int = 64,
|
||
blocksparse_head_sliding_step: int = 0,
|
||
) -> None:
|
||
num_seqs, num_heads, head_size = query.shape
|
||
gqa_ratio = num_heads // num_kv_heads
|
||
max_num_partitions = tmp_output.shape[2]
|
||
|
||
# Initialize unused slots
|
||
max_logits.fill_(float('-inf'))
|
||
exp_sums.zero_()
|
||
tmp_output.zero_()
|
||
|
||
# CCCL kernel_reduce.cuh SingleTile fast path (line ~270):
|
||
# if (num_items <= threads_per_block * items_per_thread)
|
||
# → InvokeSingleTile() — one CTA, no temp buffer, no Phase 2
|
||
# PyTorch translation: if seq_len fits in one partition, skip Phase 2 entirely.
|
||
# This avoids the partition/reshape/bmm overhead for short decode sequences.
|
||
# Qwen3.6 typical decode: seq_len grows from 1 to 100K over generation.
|
||
# Early tokens (seq_len < 1024) hit this fast path every step.
|
||
_SINGLE_TILE_THRESHOLD = _PARTITION_SIZE # sequences this short skip partitioning
|
||
|
||
for seq_idx in range(num_seqs):
|
||
seq_len = int(seq_lens[seq_idx].item())
|
||
if seq_len == 0:
|
||
output[seq_idx].zero_()
|
||
continue
|
||
|
||
num_blocks_seq = (seq_len + block_size - 1) // block_size
|
||
num_partitions = (seq_len + _PARTITION_SIZE - 1) // _PARTITION_SIZE
|
||
|
||
# ─── CCCL SingleTile fast path ───────────────────────────
|
||
# From kernel_reduce.cuh: when everything fits in one tile,
|
||
# do a single-pass attention without partition overhead.
|
||
# agent_reduce.cuh ConsumeRange → BlockReduce → done.
|
||
if num_partitions == 1:
|
||
blk_ids = block_tables[seq_idx, :num_blocks_seq]
|
||
q = query[seq_idx].float() # [H, d]
|
||
|
||
# Gather KV (same as below but no partition reshape)
|
||
k_gathered = key_cache[blk_ids]
|
||
k_flat = (k_gathered
|
||
.permute(0, 3, 1, 2, 4)
|
||
.reshape(-1, num_kv_heads, head_size))[:seq_len]
|
||
v_flat = (value_cache[blk_ids]
|
||
.permute(0, 3, 1, 2)
|
||
.reshape(-1, num_kv_heads, head_size))[:seq_len]
|
||
|
||
if k_scale != 1.0:
|
||
k_flat = k_flat.float().mul_(k_scale)
|
||
if v_scale != 1.0:
|
||
v_flat = v_flat.float().mul_(v_scale)
|
||
|
||
if gqa_ratio > 1:
|
||
k_kv = k_flat.permute(1, 2, 0).float().contiguous()
|
||
v_kv = v_flat.permute(1, 0, 2).float().contiguous()
|
||
q_grouped = q.view(num_kv_heads, gqa_ratio, 1, head_size)
|
||
scores = torch.matmul(q_grouped, k_kv.unsqueeze(1)).squeeze(2)
|
||
scores = scores.reshape(num_heads, seq_len) * scale
|
||
else:
|
||
k_t = k_flat.permute(1, 2, 0).float().contiguous()
|
||
scores = torch.bmm(q.unsqueeze(1), k_t).squeeze(1) * scale
|
||
|
||
if alibi_slopes is not None:
|
||
positions = torch.arange(seq_len, device=query.device, dtype=torch.float32)
|
||
scores = scores + alibi_slopes.unsqueeze(1) * positions.unsqueeze(0)
|
||
|
||
# Direct softmax + V weighted sum — no partition overhead
|
||
weights = torch.softmax(scores, dim=-1) # [H, seq_len]
|
||
if gqa_ratio > 1:
|
||
w_grouped = weights.view(num_kv_heads, gqa_ratio, 1, seq_len)
|
||
result = torch.matmul(w_grouped, v_kv.unsqueeze(1)).squeeze(2)
|
||
output[seq_idx] = result.reshape(num_heads, head_size).to(output.dtype)
|
||
else:
|
||
v_perm = v_flat.permute(1, 0, 2).float().contiguous()
|
||
result = torch.bmm(weights.unsqueeze(1), v_perm).squeeze(1)
|
||
output[seq_idx] = result.to(output.dtype)
|
||
|
||
# Store dummy partition values for compatibility
|
||
max_logits[seq_idx, :, 0] = scores.max(dim=-1).values
|
||
exp_sums[seq_idx, :, 0] = weights.sum(dim=-1)
|
||
tmp_output[seq_idx, :, 0, :] = output[seq_idx].float()
|
||
continue
|
||
# ─── End SingleTile fast path ────────────────────────────
|
||
|
||
# =============================================================
|
||
# Batched KV gather: ONE index_select, ONE reshape
|
||
# Pattern: avoid per-block Python loop (CCCL does this via
|
||
# block-cooperative load, we do it via batched indexing)
|
||
# =============================================================
|
||
blk_ids = block_tables[seq_idx, :num_blocks_seq]
|
||
|
||
# Key: [nblk, kv_h, d/x, blk_sz, x] → [nblk*blk_sz, kv_h, d]
|
||
k_gathered = key_cache[blk_ids]
|
||
k_flat = (k_gathered
|
||
.permute(0, 3, 1, 2, 4)
|
||
.reshape(-1, num_kv_heads, head_size))[:seq_len]
|
||
|
||
# Value: [nblk, kv_h, d, blk_sz] → [nblk*blk_sz, kv_h, d]
|
||
v_flat = (value_cache[blk_ids]
|
||
.permute(0, 3, 1, 2)
|
||
.reshape(-1, num_kv_heads, head_size))[:seq_len]
|
||
|
||
if k_scale != 1.0:
|
||
k_flat = k_flat.float().mul_(k_scale)
|
||
if v_scale != 1.0:
|
||
v_flat = v_flat.float().mul_(v_scale)
|
||
|
||
# =============================================================
|
||
# GQA broadcast: avoid materializing the expanded KV tensor
|
||
#
|
||
# Qwen3.6: H=24, kv_h=4, gqa_ratio=6, head_dim=256
|
||
# Old: expand kv_h→H then contiguous → allocates seq_len×H×d (1.2GB at 100K)
|
||
# New: reshape Q as [kv_h, gqa, 1, d], K as [kv_h, 1, d, seq_len]
|
||
# → bmm with broadcasting → [kv_h, gqa, 1, seq_len]
|
||
# → reshape to [H, seq_len]
|
||
# Saves: gqa_ratio × memory (6x for Qwen3.6 = 1GB per decode step)
|
||
# =============================================================
|
||
q = query[seq_idx].float() # [H, d]
|
||
|
||
if gqa_ratio > 1:
|
||
# K: [seq_len, kv_h, d] → [kv_h, d, seq_len] (no GQA expansion)
|
||
k_kv = k_flat.permute(1, 2, 0).float().contiguous() # [kv_h, d, seq_len]
|
||
v_kv = v_flat.permute(1, 0, 2).float().contiguous() # [kv_h, seq_len, d]
|
||
|
||
# Q: [H, d] → [kv_h, gqa, 1, d]
|
||
q_grouped = q.view(num_kv_heads, gqa_ratio, 1, head_size)
|
||
|
||
# Scores: [kv_h, gqa, 1, d] @ [kv_h, 1, d, seq_len] → [kv_h, gqa, 1, seq_len]
|
||
scores_all = torch.matmul(q_grouped, k_kv.unsqueeze(1)).squeeze(2) # [kv_h, gqa, seq_len]
|
||
scores_all = scores_all.reshape(num_heads, seq_len) * scale # [H, seq_len]
|
||
else:
|
||
k_t = k_flat.permute(1, 2, 0).float().contiguous() # [H, d, seq_len]
|
||
scores_all = torch.bmm(q.unsqueeze(1), k_t).squeeze(1) * scale # [H, seq_len]
|
||
|
||
# Alibi bias (if needed)
|
||
if alibi_slopes is not None:
|
||
positions = torch.arange(seq_len, device=query.device, dtype=torch.float32)
|
||
scores_all = scores_all + alibi_slopes.unsqueeze(1) * positions.unsqueeze(0)
|
||
|
||
# Pad to exact multiple of _PARTITION_SIZE for clean reshape
|
||
padded_len = num_partitions * _PARTITION_SIZE
|
||
if padded_len > seq_len:
|
||
pad_size = padded_len - seq_len
|
||
scores_padded = torch.full(
|
||
(num_heads, padded_len), float('-inf'),
|
||
dtype=scores_all.dtype, device=scores_all.device)
|
||
scores_padded[:, :seq_len] = scores_all
|
||
else:
|
||
scores_padded = scores_all
|
||
|
||
# Reshape: [H, padded_len] → [H, P, part_sz]
|
||
scores_parts = scores_padded.view(num_heads, num_partitions, _PARTITION_SIZE)
|
||
|
||
# Per-partition online softmax (vectorized over H and P simultaneously)
|
||
# Pattern from CCCL summary_statistics: compute (max, sum) in one pass
|
||
part_max = scores_parts.max(dim=-1).values # [H, P]
|
||
scores_exp = torch.exp(scores_parts - part_max.unsqueeze(-1)) # [H, P, part_sz]
|
||
part_sum = scores_exp.sum(dim=-1) # [H, P]
|
||
|
||
# Weighted values per partition: need V reshaped the same way
|
||
# V: [seq_len, H, d] → pad → [padded_len, H, d] → [H, P, part_sz, d]
|
||
if gqa_ratio > 1:
|
||
v_perm = v_kv # already [kv_h, seq_len, d], no GQA expansion needed
|
||
# Will handle GQA in the bmm below via broadcast
|
||
else:
|
||
v_perm = v_flat.permute(1, 0, 2).float().contiguous() # [H, seq_len, d]
|
||
# Weighted V sum per partition
|
||
# NOTE: v_perm shape differs by GQA mode:
|
||
# GQA: v_perm = v_kv = [kv_h, seq_len, d]
|
||
# No GQA: v_perm = [H, seq_len, d]
|
||
# scores_exp: [H, P, part_sz] → [kv_h, gqa, P, part_sz]
|
||
# v_perm: [kv_h, seq_len, d] → [kv_h, P, part_sz, d]
|
||
if gqa_ratio > 1:
|
||
se_grouped = scores_exp.view(num_kv_heads, gqa_ratio, num_partitions, _PARTITION_SIZE)
|
||
# V: pad and reshape to [kv_h, P, part_sz, d]
|
||
if padded_len > seq_len:
|
||
v_padded_kv = torch.zeros(
|
||
(num_kv_heads, padded_len, head_size),
|
||
dtype=v_kv.dtype, device=v_kv.device)
|
||
v_padded_kv[:, :seq_len, :] = v_kv
|
||
else:
|
||
v_padded_kv = v_kv
|
||
v_parts_kv = v_padded_kv.view(num_kv_heads, num_partitions, _PARTITION_SIZE, head_size)
|
||
# Broadcast: [kv_h, gqa, P, 1, part_sz] @ [kv_h, 1, P, part_sz, d]
|
||
# → [kv_h, gqa, P, 1, d]
|
||
part_out_grouped = torch.matmul(
|
||
se_grouped.unsqueeze(3), # [kv_h, gqa, P, 1, part_sz]
|
||
v_parts_kv.unsqueeze(1) # [kv_h, 1, P, part_sz, d]
|
||
).squeeze(3) # [kv_h, gqa, P, d]
|
||
part_out = part_out_grouped.reshape(num_heads, num_partitions, head_size)
|
||
else:
|
||
# Non-GQA: v_perm is [H, seq_len, d], pad and reshape normally
|
||
if padded_len > seq_len:
|
||
v_padded = torch.zeros(
|
||
(num_heads, padded_len, head_size),
|
||
dtype=v_perm.dtype, device=v_perm.device)
|
||
v_padded[:, :seq_len, :] = v_perm
|
||
else:
|
||
v_padded = v_perm
|
||
v_parts = v_padded.view(num_heads, num_partitions, _PARTITION_SIZE, head_size)
|
||
HP = num_heads * num_partitions
|
||
scores_exp_flat = scores_exp.reshape(HP, 1, _PARTITION_SIZE)
|
||
v_parts_flat = v_parts.reshape(HP, _PARTITION_SIZE, head_size)
|
||
part_out_flat = torch.bmm(scores_exp_flat, v_parts_flat) # [HP, 1, d]
|
||
part_out = part_out_flat.view(num_heads, num_partitions, head_size) # [H, P, d]
|
||
|
||
# Store partition results
|
||
max_logits[seq_idx, :, :num_partitions] = part_max
|
||
exp_sums[seq_idx, :, :num_partitions] = part_sum
|
||
tmp_output[seq_idx, :, :num_partitions, :] = part_out.to(tmp_output.dtype)
|
||
|
||
# =============================================================
|
||
# Phase 2: Cross-partition reduction (CCCL binary_op pattern)
|
||
#
|
||
# CCCL kernel_reduce.cuh insight: when grid_size fits in a single
|
||
# tile (num_partitions <= threads * items_per_thread), the reduce
|
||
# uses SingleTile path — one CTA, no temp buffer, no pass 2 kernel.
|
||
#
|
||
# For BI-V100 with 98 partitions (100K tokens / 1024 partition_size):
|
||
# SingleTile threshold = 512 * 24 = 12288 >> 98 → always SingleTile
|
||
# This means Phase 2 is never the bottleneck.
|
||
#
|
||
# CCCL single_pass_scan_operators.cuh insight: delay() has a
|
||
# GridThreshold=500 gate. BI-V100 scan grids are always < 500 blocks,
|
||
# so ALL delay strategies (no_delay, fixed_delay, exponential_backon)
|
||
# collapse to __threadfence_block(). Delay tuning is irrelevant here.
|
||
#
|
||
# Phase 2 follows summary_statistics.cu binary_op: combine
|
||
# (max_a, sum_a, out_a) ⊕ (max_b, sum_b, out_b) via log-sum-exp.
|
||
# Fully vectorized — no loop over partitions.
|
||
# =============================================================
|
||
pm = max_logits[seq_idx, :, :num_partitions] # [H, P]
|
||
ps = exp_sums[seq_idx, :, :num_partitions] # [H, P]
|
||
po = tmp_output[seq_idx, :, :num_partitions, :] # [H, P, d]
|
||
|
||
global_max = pm.max(dim=-1).values # [H]
|
||
rescale = torch.exp(pm - global_max.unsqueeze(-1)) * ps # [H, P]
|
||
total = rescale.sum(dim=-1, keepdim=True) # [H, 1]
|
||
|
||
# CCCL norm.cu principle: fuse transform with reduce to minimize traversals.
|
||
# Instead of: weights = rescale/total; final = bmm(weights, po)
|
||
# Do: final = bmm(rescale, po) / total
|
||
# Saves one element-wise division kernel launch (rescale/total → H*P elements).
|
||
# The division moves to the output (H*d elements, typically smaller than H*P).
|
||
# [H, 1, P] @ [H, P, d] → [H, 1, d] → [H, d]
|
||
final = torch.bmm(rescale.unsqueeze(1), po.float()).squeeze(1) / total # [H, d]
|
||
output[seq_idx] = final.to(output.dtype)
|