[Infer] Revise and Adapt Triton Kernels for Spec-Dec (#5401)

* [Infer/Fix] Fix Dependency in test - RMSNorm kernel (#5399)

fix dependency in pytest

* resolve conflicts for revising flash-attn

* adapt kv cache copy kernel for spec-dec

* fix seqlen-n kvcache copy kernel/tests

* test kvcache copy - use torch.equal

* add assertions

* (trivial) comment out

colossalai/kernel/triton/__init__.py

@@ -12,7 +12,7 @@
     from .flash_decoding import flash_decoding_attention
     from .fused_rotary_embedding import fused_rotary_embedding
     from .gptq_triton import gptq_fused_linear_triton
-    from .kvcache_copy import copy_kv_to_blocked_cache
+    from .kvcache_copy import copy_k_to_blocked_cache, copy_kv_to_blocked_cache
     from .no_pad_rotary_embedding import decoding_fused_rotary_embedding, rotary_embedding
     from .rms_layernorm import rms_layernorm
     from .rotary_cache_copy import get_xine_cache
@@ -21,6 +21,7 @@
     __all__ = [
         "context_attention_unpadded",
         "flash_decoding_attention",
+        "copy_k_to_blocked_cache",
         "copy_kv_to_blocked_cache",
         "softmax",
         "rms_layernorm",

colossalai/kernel/triton/flash_decoding.py

@@ -9,13 +9,14 @@
 # Triton 2.1.0
 @triton.jit
 def _flash_decoding_fwd_kernel(
-    Q,  # [batch_size, head_num, q_len(1), head_dim]
+    Q,  # [batch_size * q_len, head_num, head_dim]
     KCache,  # [num_blocks, num_kv_heads, block_size, head_dim]
     VCache,  # [num_blocks, num_kv_heads, block_size, head_dim]
     block_tables,  # [batch_size, max_blocks_per_sequence]
-    mid_o,  # [batch_size, head_num, kv_split_num, head_dim]
-    mid_o_lse,  # [batch_size, head_num, kv_split_num]
+    mid_o,  # [batch_size * q_len, head_num, kv_split_num, head_dim]
+    mid_o_lse,  # [batch_size * q_len, head_num, kv_split_num]
     kv_seq_len,  # [batch_size]
+    q_len,
     batch_size,
     stride_qt,
     stride_qh,
@@ -39,44 +40,37 @@ def _flash_decoding_fwd_kernel(
     BLOCK_SIZE: tl.constexpr,
     HEAD_DIM: tl.constexpr,
 ):
-    cur_seq_idx = tl.program_id(0)
+    cur_token_idx = tl.program_id(0)
+    cur_seq_idx = cur_token_idx // q_len
     if cur_seq_idx >= batch_size:
         return
     cur_head_idx = tl.program_id(1)
     block_start_kv = tl.program_id(2)  # for splitting k/v
 
-    cur_kv_head_idx = cur_head_idx // KV_GROUPS
-    offsets_dmodel = tl.arange(0, HEAD_DIM)
-
     # NOTE It requires BLOCK_KV and BLOCK_SIZE to be the same
     # TODO might want to replace with BLOCK_KV % BLOCK_SIZE == 0 (optimize BLOCK_KV as multiple of BLOCK_SIZE)
     #      and then support calculating multiple kv cache blocks on an instance
     tl.static_assert(BLOCK_KV == BLOCK_SIZE)
-
-    # get the current (kv) sequence length from provided context lengths tensor
+    # get the current (kv) sequence length
     cur_kv_seq_len = tl.load(kv_seq_len + cur_seq_idx)
+    if block_start_kv * BLOCK_KV >= cur_kv_seq_len:
+        return
 
-    offsets_q = cur_seq_idx * stride_qt + cur_head_idx * stride_qh + offsets_dmodel * stride_qd
+    offsets_dmodel = tl.arange(0, HEAD_DIM)
+    offsets_q = cur_token_idx * stride_qt + cur_head_idx * stride_qh + offsets_dmodel * stride_qd
     q = tl.load(Q + offsets_q)
-
     # block table for the current sequence
     block_table_ptr = block_tables + cur_seq_idx * stride_bts
-
-    # actually current block table current block start idx
     # cur_bt_start_idx = block_start_kv * (BLOCK_KV // BLOCK_SIZE)
-    cur_bt_start_idx = block_start_kv
-    cur_block_id = tl.load(block_table_ptr + cur_bt_start_idx * stride_btb)
-
-    if block_start_kv * BLOCK_KV >= cur_kv_seq_len:
-        return
-
+    # cur_block_id = tl.load(block_table_ptr + cur_bt_start_idx * stride_btb)
+    cur_block_id = tl.load(block_table_ptr + block_start_kv * stride_btb)
     cur_occupied_size = tl.where(
         (block_start_kv + 1) * BLOCK_SIZE <= cur_kv_seq_len, BLOCK_SIZE, cur_kv_seq_len - block_start_kv * BLOCK_SIZE
     )
     tl.device_assert(cur_occupied_size >= 0)
 
+    cur_kv_head_idx = cur_head_idx // KV_GROUPS
     offset_kvcache = cur_block_id * stride_cacheb + cur_kv_head_idx * stride_cacheh
-
     K_block_ptr = tl.make_block_ptr(
         base=KCache + offset_kvcache,
         shape=(cur_occupied_size, HEAD_DIM),
@@ -115,14 +109,14 @@ def _flash_decoding_fwd_kernel(
     acc = acc / l
 
     offsets_mid_o = (
-        cur_seq_idx * stride_mid_ot
+        cur_token_idx * stride_mid_ot
         + cur_head_idx * stride_mid_oh
         + block_start_kv * stride_mid_ob
         + offsets_dmodel * stride_mid_od
     )
     tl.store(mid_o + offsets_mid_o, acc)
     offsets_mid_o_lse = (
-        cur_seq_idx * stride_mid_o_lset + cur_head_idx * stride_mid_o_lseh + block_start_kv * stride_mid_o_lseb
+        cur_token_idx * stride_mid_o_lset + cur_head_idx * stride_mid_o_lseh + block_start_kv * stride_mid_o_lseb
     )
     # logsumexp L^(j) = m^(j) + log(l^(j))
     tl.store(mid_o_lse + offsets_mid_o_lse, m + tl.log(l))
@@ -135,6 +129,7 @@ def _flash_decoding_fwd_reduce_kernel(
     mid_o_lse,  # [batch_size, head_num, kv_split_num]
     O,  # [batch_size, num_heads, head_dim] or [batch_size, 1, num_heads, head_dim]
     kv_seq_len,
+    q_len,
     batch_size,
     stride_mid_ot,
     stride_mid_oh,
@@ -149,7 +144,8 @@ def _flash_decoding_fwd_reduce_kernel(
     BLOCK_KV: tl.constexpr,
     HEAD_DIM: tl.constexpr,
 ):
-    cur_seq_idx = tl.program_id(0)
+    cur_token_idx = tl.program_id(0)
+    cur_seq_idx = cur_token_idx // q_len
     if cur_seq_idx >= batch_size:
         return
     cur_head_idx = tl.program_id(1)
@@ -164,8 +160,8 @@ def _flash_decoding_fwd_reduce_kernel(
     l = 0.0  # sum exp
     acc = tl.zeros([HEAD_DIM], dtype=tl.float32)
 
-    offsets_mid_o = cur_seq_idx * stride_mid_ot + cur_head_idx * stride_mid_oh + offsets_dmodel
-    offset_mid_lse = cur_seq_idx * stride_o_lset + cur_head_idx * stride_o_lseh
+    offsets_mid_o = cur_token_idx * stride_mid_ot + cur_head_idx * stride_mid_oh + offsets_dmodel
+    offset_mid_lse = cur_token_idx * stride_o_lset + cur_head_idx * stride_o_lseh
     for block_i in range(0, kv_split_num, 1):
         mid_o_block = tl.load(mid_o + offsets_mid_o + block_i * stride_mid_ob)
         lse = tl.load(mid_o_lse + offset_mid_lse + block_i * stride_o_lseb)
@@ -179,7 +175,7 @@ def _flash_decoding_fwd_reduce_kernel(
         m_i = m_ij
 
     acc = acc / l
-    offsets_O = cur_seq_idx * stride_ot + cur_head_idx * stride_oh + offsets_dmodel
+    offsets_O = cur_token_idx * stride_ot + cur_head_idx * stride_oh + offsets_dmodel
     tl.store(O + offsets_O, acc.to(O.type.element_ty))
     return
 
@@ -199,32 +195,40 @@ def flash_decoding_attention(
     mid_output_lse: torch.Tensor = None,
     sm_scale: int = None,
     kv_group_num: int = 1,
+    q_len: int = 1,
 ):
     """
     Flash decoding implemented with a blocked KV Cache (PagedAttention) during decoding stage.
 
     Args:
-        q (torch.Tensor):       [bsz, num_heads, head_dim]
+        q (torch.Tensor): [bsz * q_len, num_heads, head_dim]
+            q_len > 1 only for verification process in speculative-decoding.
         k_cache (torch.Tensor): [num_blocks, num_kv_heads, block_size, head_dim]
         v_cache (torch.Tensor): [num_blocks, num_kv_heads, block_size, head_dim]
         kv_seq_len (torch.Tensor): [batch_size]
             records the (kv) sequence lengths incorporating past kv sequence lengths.
         block_tables (torch.Tensor): [batch_size, max_blocks_per_sequence]
         max_seq_len_in_batch (int): Maximum sequence length in the batch.
         output (torch.Tensor):  [bsz, num_heads * head_dim]
-        mid_output (torch.Tensor): [ max_bsz , num_heads, kv_max_split_num, head_dim]
+        mid_output (torch.Tensor): [max_bsz * q_len, num_heads, kv_max_split_num, head_dim]
             Intermediate output tensor. `max_bsz` should be greater than or equal to `bsz`.
-        mid_output_lse (torch.Tensor): [ max_bsz , num_heads, kv_max_split_num]
+            q_len > 1 only for verification process in speculative-decoding.
+        mid_output_lse (torch.Tensor): [max_bsz * q_len, num_heads, kv_max_split_num]
             Log-sum-exp of intermediate output. `max_bsz` should be greater than or equal to `bsz`.
+            q_len > 1 only for verification process in speculative-decoding.
         block_size (int): Size of each block in the blocked key/value cache.
         num_kv_group (int, optional): Number of key/value groups. Defaults to 1.
+        q_length (int): Query length. Use for speculative decoding when `q_length` > 1 (i.e. the last n tokens).
+            Defaults to 1.
 
     Returns:
-        Output tensor with shape [bsz, num_heads * head_dim]
+        Output tensor with shape [bsz * q_len, num_heads * head_dim]
     """
     q = q.squeeze() if q.dim() == 4 else q
     assert q.dim() == 3, f"Incompatible q dim: {q.dim()}"
-    bsz, num_heads, head_dim = q.shape
+    n_tokens, num_heads, head_dim = q.shape
+    assert n_tokens % q_len == 0, "Invalid q_len"
+    bsz = n_tokens // q_len
 
     assert head_dim in {32, 64, 128, 256}
     assert kv_seq_len.shape[0] == block_tables.shape[0] == bsz, (
@@ -247,22 +251,31 @@ def flash_decoding_attention(
     max_seq_len_in_batch = kv_seq_len.max().item() if max_seq_len_in_batch is None else max_seq_len_in_batch
     # For compatibility (TODO revise modeling in future)
     kv_max_split_num = (max_seq_len_in_batch + BLOCK_KV - 1) // BLOCK_KV
-    mid_output = (
-        torch.zeros(size=(bsz, num_heads, kv_max_split_num, head_dim), dtype=torch.float32, device=q.device)
-        if mid_output is None
-        else mid_output
-    )
-    mid_output_lse = (
-        torch.zeros(size=(bsz, num_heads, kv_max_split_num), dtype=torch.float32, device=q.device)
-        if mid_output_lse is None
-        else mid_output_lse
-    )
+
+    if mid_output is None:
+        mid_output = torch.empty(
+            (bsz * q_len, num_heads, kv_max_split_num, head_dim), dtype=torch.float32, device=q.device
+        )
+    if mid_output_lse is None:
+        mid_output_lse = torch.empty((bsz * q_len, num_heads, kv_max_split_num), dtype=torch.float32, device=q.device)
+    if output is None:
+        # A hack to prevent `view` operation in modeling
+        output = torch.empty((bsz * q_len, num_heads * head_dim), dtype=q.dtype, device=q.device)
+
+    assert (
+        mid_output.size(2) == mid_output_lse.size(2) >= kv_max_split_num
+    ), "Incompatible kv split number of intermediate output tensors"
+    assert (
+        mid_output.size(0) == mid_output_lse.size(0) >= output.size(0) == n_tokens
+    ), f"Incompatible first dimension of output tensors"
 
     # NOTE use `triton.next_power_of_2` here to utilize the cache mechanism of triton
     # To optimize, revise batching/scheduling to batch 2^n sequences in a batch (preferred)
-    grid = (triton.next_power_of_2(bsz), num_heads, triton.cdiv(triton.next_power_of_2(max_seq_len_in_batch), BLOCK_KV))
-    output = torch.empty((bsz, num_heads * head_dim), dtype=q.dtype, device=q.device) if output is None else output
-
+    grid = (
+        triton.next_power_of_2(bsz * q_len),
+        num_heads,
+        triton.cdiv(triton.next_power_of_2(max_seq_len_in_batch), BLOCK_KV),
+    )
     _flash_decoding_fwd_kernel[grid](
         q,
         k_cache,
@@ -271,6 +284,7 @@ def flash_decoding_attention(
         mid_output,
         mid_output_lse,
         kv_seq_len,
+        q_len,
         bsz,
         q.stride(0),
         q.stride(1),
@@ -295,13 +309,13 @@ def flash_decoding_attention(
         HEAD_DIM=head_dim,
     )
 
-    grid = (triton.next_power_of_2(bsz), num_heads)
-
+    grid = (triton.next_power_of_2(bsz * q_len), num_heads)
     _flash_decoding_fwd_reduce_kernel[grid](
         mid_output,
         mid_output_lse,
         output,
         kv_seq_len,
+        q_len,
         bsz,
         mid_output.stride(0),
         mid_output.stride(1),

colossalai/kernel/triton/kvcache_copy.py

@@ -3,6 +3,50 @@
 import triton.language as tl
 
 
+# Triton 2.1.0
+@triton.jit
+def _copy_to_kcache_seqlen_n_kernel(
+    KV,  # K or V
+    KVCache,  # KCache or VCache
+    BLOCK_TABLES,
+    context_lengths,
+    stride_kt,
+    stride_kh,
+    stride_kd,
+    stride_cacheb,
+    stride_cacheh,
+    stride_cachebs,
+    stride_cached,
+    stride_bts,
+    stride_btb,
+    block_size,
+    n,
+    HEAD_DIM: tl.constexpr,
+):
+    cur_token_idx = tl.program_id(0)
+    cur_seq_idx = cur_token_idx // n
+    cur_token_shift = cur_token_idx - (n * (cur_seq_idx + 1))
+    # cur_token_shift = cur_token_idx - n * cur_seq_idx
+    cur_kv_head_idx = tl.program_id(1)
+
+    past_kv_seq_len = tl.load(context_lengths + cur_seq_idx) + cur_token_shift
+    last_bt_block_idx = past_kv_seq_len // block_size
+    block_table_ptr = BLOCK_TABLES + cur_seq_idx * stride_bts
+    block_id = tl.load(block_table_ptr + last_bt_block_idx * stride_btb)
+    offset_last_block = past_kv_seq_len % block_size
+    offsets_dmodel = tl.arange(0, HEAD_DIM)
+    offsets_kv = cur_token_idx * stride_kt + cur_kv_head_idx * stride_kh + offsets_dmodel * stride_kd
+    kv = tl.load(KV + offsets_kv)
+    offsets_kvcache = (
+        block_id * stride_cacheb
+        + cur_kv_head_idx * stride_cacheh
+        + offset_last_block * stride_cachebs
+        + offsets_dmodel * stride_cached
+    )
+    tl.store(KVCache + offsets_kvcache, kv)
+    return
+
+
 # Triton 2.1.0
 @triton.jit
 def _copy_to_kvcache_seqlen1_kernel(
@@ -40,10 +84,11 @@ def _copy_to_kvcache_seqlen1_kernel(
     block_id = tl.load(block_table_ptr + last_bt_block_idx * stride_btb)
     offsets_in_last_block = past_kv_seq_len % block_size
     offsets_dmodel = tl.arange(0, HEAD_DIM)
-    offsets_kv = cur_seq_idx * stride_kt + cur_kv_head_idx * stride_kh + offsets_dmodel * stride_kd
+    offsets_k = cur_seq_idx * stride_kt + cur_kv_head_idx * stride_kh + offsets_dmodel * stride_kd
+    offsets_v = cur_seq_idx * stride_vt + cur_kv_head_idx * stride_vh + offsets_dmodel * stride_vd
 
-    k = tl.load(K + offsets_kv)
-    v = tl.load(V + offsets_kv)
+    k = tl.load(K + offsets_k)
+    v = tl.load(V + offsets_v)
 
     offsets_kcache = (
         block_id * stride_cachekb
@@ -63,6 +108,64 @@ def _copy_to_kvcache_seqlen1_kernel(
     return
 
 
+def copy_k_to_blocked_cache(
+    k: torch.Tensor, k_cache: torch.Tensor, kv_lengths: torch.Tensor, block_tables: torch.Tensor, n: int = 1
+):
+    """
+    Copy keys or values to the blocked key/value cache during decoding stage.
+
+    Args:
+        k (torch.Tensor): [bsz, 1, num_kv_heads, head_dim]/[bsz, num_kv_heads, head_dim] - Keys or values during decoding with seq len 1.
+            [bsz * n, num_kv_heads, head_dim] - Keys or values with seq len n
+        k_cache (torch.Tensor): [num_blocks, num_kv_heads, block_size, head_dim] - Blocked key or value cache.
+        kv_lengths (torch.Tensor): [bsz] - Past key/value sequence lengths plus current sequence length for each sequence.
+        block_tables (torch.Tensor): [bsz, max_blocks_per_sequence] - Block tables for each sequence.
+        n (int): Number of tokens to copy for each sequence. Default to 1.
+    """
+    assert k.size(-1) == k_cache.size(-1), "Incompatible head dim"
+    assert k.dtype == k_cache.dtype, "Expected consistent dtype for tensor and cache."
+
+    k = k.reshape(-1, k.size(-2), k.size(-1)) if k.dim() == 4 else k
+    assert k.dim() == 3, f"Invalid k dim {k.dim()}"
+    bsz, num_kv_heads, head_dim = k.shape
+    # NOTE when n > 1, the shape of k is [bsz * n, num_kv_heads, head_dim]
+    if n > 1:
+        assert bsz % n == 0, "Each sequence should have the same number of tokens to be copied"
+        bsz = bsz // n
+
+    assert kv_lengths.shape[0] == block_tables.shape[0] == bsz, (
+        f"Got incompatible batch size (number of seqs):\n"
+        f"  Past kv sequence lengths bsz {kv_lengths.shape[0]}; "
+        f" block tables bsz {block_tables.shape[0]}, input k batch size {bsz}"
+    )
+
+    # Modify if the shape of kv cahce is changed.
+    block_size = k_cache.size(-2)
+
+    num_warps = 8 if head_dim > 128 else 4
+
+    grid = (bsz * n, num_kv_heads)
+    _copy_to_kcache_seqlen_n_kernel[grid](
+        k,
+        k_cache,
+        block_tables,
+        kv_lengths,
+        k.stride(0),
+        k.stride(1),
+        k.stride(2),
+        k_cache.stride(0),
+        k_cache.stride(1),
+        k_cache.stride(2),
+        k_cache.stride(3),
+        block_tables.stride(0),
+        block_tables.stride(1),
+        block_size,
+        n=n,
+        HEAD_DIM=head_dim,
+        num_warps=num_warps,
+    )
+
+
 def copy_kv_to_blocked_cache(
     k: torch.Tensor,
     v: torch.Tensor,