# Compressor ## 产品支持情况 | 产品 | 是否支持 | | ------------------------------------------------------------ | :------: | |Ascend 950PR/Ascend 950DT| √ | |Atlas A3 训练系列产品/Atlas A3 推理系列产品| √ | |Atlas A2 训练系列产品/Atlas A2 推理系列产品| × | |Atlas 200I/500 A2 推理产品| × | |Atlas 推理系列加速卡产品| × | |Atlas 训练系列产品| × | ## 功能说明 - API功能：Compressor是推理场景下SAS和QLI的前处理算子，用于将每4或128个token的KV cache压缩成一个，然后每个token与这些压缩的KV cache进行DSA计算。在长序列的情况下，Compressor可以有效地减少计算开销。 - 计算公式： 压缩阶段： 1. 计算矩阵乘法： - C4A: $\left[kv\_state^a, score\_state^a\right] = X @ \left[W^{aKV}, W^{aGate}\right], \left[kv\_state^b, score\_state^b\right] = X @ \left[W^{bKV}, W^{bGate}\right];$ - C128A: $\left[kv\_state, score\_state\right] = X @ \left[W^{KV}, W^{Gate}\right]$ 2. 计算分组加法： - C4A: $score\_state_i^\prime = \left[score\_state_{\left[4(i-1)+1:4i,:\right]}^a; score\_state_{\left[4i+1:4(i+1),:\right]}^b\right] + Ape,~i=1,2,\cdots, \frac{s}{4};$ - C128A: $score\_state_i^\prime = score\_state_{\left[128(i-1)+1:128i,:\right]} + Ape,~i=1,2,\cdots, \frac{s}{128};$ 3. 计算分组Softmax： - C4A: $S_i^\prime = softmax(score\_state_i^\prime),~i=1,2,\cdots, \frac{s}{4};$ - C128A: $S_i^\prime = softmax(score\_state_i^\prime),~i=1,2,\cdots, \frac{s}{128};$ 4. 计算Hadamard乘积： - C4A: $(S_H)_i = S_i^\prime \odot \left[kv\_state^a_{\left[4(i-1)+1:4i,:\right]} ; kv\_state^b_{\left[4i+1:4(i+1),:\right]}\right],~i=1,2,\cdots, \frac{s}{4};$ - C128A: $S_H = S_i^\prime \odot kv\_state;$ 5. 沿着压缩轴分组求和： - C4A: $C_{i}^{\text{Comp}} = \left[1\right]_{1\times8} @ (S_H)_i, ~i=1,2,\cdots, \frac{s}{4};$ - C128A: $C_{i}^{\text{Comp}} = \left[1\right]_{1\times128} @ (S_H)_i, ~i=1,2,\cdots, \frac{s}{128};$ 后处理阶段： 6. 计算RMSNorm： - $\text{RMS}(C^{\text{Comp}}) = \sqrt{\frac{1}{N} \sum_{i=j* N}^{(j+1)* N} {(C_{i}^{\text{Comp}})}^{\text{2}} + norm\_eps} ,N=head\_dim, ~j=1,2,\cdots, \frac{s}{cmp\_ratio}$ - $\text{RmsNorm}(C^{\text{Comp}}) = norm\_weight \cdot \frac{C_{i}^{\text{Comp}}}{\text{RMS}(C^{\text{Comp}})}$ 7. 计算Rope； - 主要计算过程为： 1. 将输入$X$与$W^{KV}$做Matmul运算得到$kv\_state$，将输入$X$与$W^{Gate}$做Matmul运算后再与$Ape$做Add运算得到$score\_state$，$kv\_state$与$score\_state$根据输入的start_pos及cu_seqlens完成更新。 2. 在coff为2的情况下对$kv\_state$和$score\_state$进行数据重排。 3. 对$score\_state$进行softmax运算将softmax结果与$kv\_state$做Mul计算，后进行ReduceSum运算。 4. 根据输入数据norm_weight、rope_sin、rope_cos，进行RMSNorm和Rope运算，得到$cmp\_kv$结果输出。 ## 参数说明 | 参数名 | 输入/输出/属性 | 描述 | 数据类型 | 数据格式 | |----------------------------|-----------|----------------------------------------------------------------------|----------------|------------| | x | 输入 | 公式中的$X$，表示原始不经压缩的数据。 | FLOAT16、BFLOAT16 | ND | | wkv | 输入 | 公式中的$W^{KV}$，表示kv压缩权重。 | FLOAT16、BFLOAT16 | ND | | wgate | 输入 | 公式中的$W^{Gate}$，表示gate压缩权重。 | FLOAT16、BFLOAT16 | ND | | kv_state | 输入 | 公式中的$kv\_state$，表示kv\_state的历史数据。 | FLOAT32 | ND | | score_state | 输入 | 公式中的$score\_state$，表示score\_state中的历史数据。 | FLOAT32 | ND | | ape | 输入 | 公式中的$Ape$，表示positional biases。 | FLOAT32 | ND | | norm\_weight | 输入 | 表示计算RmsNorm时的权重系数。 | FLOAT16、BFLOAT16 | ND | | rope\_sin | 输入 | 表示Rope计算时sin的权重系数。 | FLOAT16、BFLOAT16 | ND | | rope\_cos | 输入 | 表示Rope计算时cos的权重系数。 | FLOAT16、BFLOAT16 | ND | | rope\_head\_dim | 属性 | 表示rope_cos和rope_sin的hidden层最小单元大小，当前仅支持64。 | INT32 | - | | cmp\_ratio | 属性 | 用于稀疏计算，表示数据压缩率。 | INT32 | - | | kv\_block\_table | 可选输入 | 表示kv\_state存储使用的block映射表。当其中元素的值为0时，表示当前位置无需进行更新kv\_state操作。 | INT32 | ND | | score\_block\_table | 可选输入 | 表示score\_state存储使用的block映射表。当其中元素的值为0时，表示当前位置无需进行更新score\_state操作。 | INT32 | ND | | cu\_seqlens | 可选输入 | 表示不同Batch中的有效token数。 | INT32 | ND | | seqused | 可选输入 | 表示不同Batch中实际参与压缩的token数，如果指定为None时，表示和每个Batch上的Sequence Length长度相同。 | INT32 | ND | | start\_pos | 可选输入 | 表示计算起始位置。 | INT32 | ND | | coff | 可选属性 | 默认值1，支持1/2。当coff=1时，无需进行overlap数据重排。当coff=2时，需要进行overlap数据重排。 | INT32 | - | | norm\_eps | 可选属性 | 表示RmsNorm计算的权重系数。默认值1e-6。 | FLOAT32 | - | | rotary\_mode | 可选属性 | 表示Rop计算的模式。默认值1，支持1/2。rotary\_mode为1时，代表half模式。rotary\_mode为2时，代表interleave模式。 | INT32 | - | | enabled\_grad | 可选属性 | 训练场景使用，表示是否参与反向更新。默认值false，支持false/true。**目前暂不支持输入true**。 | BOOL | - | | cmp\_kv | 输出 | 表示压缩后的数据。 | FLOAT16、BFLOAT16 | ND | | wkv\_proj | 可选输出 | 训练反向使用，表示wkv权重Matmul的计算结果，**目前暂不支持返回wkv\_proj**。 | FLOAT16、BFLOAT16 | ND | | softmax\_res | 可选输出 | 训练反向使用，表示Softmax计算结果，**目前暂不支持返回softmax\_res**。 | FLOAT16、BFLOAT16 | ND | | norm\_x | 可选输出 | 训练反向使用，表示Rms计算的输入，**目前暂不支持返回norm\_x**。 | FLOAT16、BFLOAT16 | ND | | norm\_rstd | 可选输出 | 训练反向使用，表示Rms计算的中间结果，**目前暂不支持返回norm\_rstd**。 | FLOAT16、BFLOAT16 | ND | ## 约束说明 - x参数维度含义：B（Batch Size）表示输入样本批量大小、S（Sequence Length）表示输入样本序列长度、H（Head Size）表示hidden层的大小、D（Head Dim）表示hidden层的最小单元大小、T表示所有Batch输入样本序列长度的累加和。 - 输入shape限制： - wkv支持输入shape[coff* D,H] - wgate支持输入shape[coff* D,H] - kv\_state、score\_state支持输入shape[block_num,block_size,coff* D]，要求block_num>0。 - ape支持输入shape[cmp_ratio,coff* D] - norm\_weight支持输入shape[D,] - start\_pos支持输入shape[B,] - 若x的维度采用BS合轴，即x的输入shape为[T,H] - rope_sin、rope_cos要求输入shape为[min(T,T//cmp_ratio+B),rope_head_dim]。 - cu\_seqlens输入shape必须为[B+1,]。该参数中每个元素的值表示当前batch与之前所有batch的token数总和，即前缀和，因此后一个元素的值必须大于等于前一个元素的值，且第一位必须位0。 - seqused，支持输入shape[B,]，要求每个Batch的有效token数要求小于等于对应Sequence Length长度，即seqused[n] <= cu\_seqlens[n+1] - cu\_seqlens[n]，且不小于0。 - kv\_block\_table、score\_block\_table支持输入shape[B,ceil(Smax/block_size)]。Smax为每个Batch中最大的Sequence Length，即Smax=max(start\_pos)+max(cu\_seqlens[n+1] - cu\_seqlens[n])。 - cmp\_kv，输出shape为[min(T,T//cmp_ratio+B),D]：compressed_tokens + compressed_tokens + ... + compressed_tokens + pad。 - wkv\_proj，输出shape为[T,coff* D]。 - norm\_x，输出shape为[min(T,T//cmp_ratio+B),D]。 - norm\_rstd，输出shape为[min(T,T//cmp_ratio+B)]。 - 若x的维度不采用BS合轴，即x的输入shape为[B,S,H] - rope_sin、rope_cos要求输入shape为[B,ceil(S/cmp_ratio),rope_head_dim]。 - cu\_seqlens，参数必须为空。 - seqused，支持输入shape[B,]，要求每个Batch的有效token数要求小于等于对应Sequence Length长度，即要求seqused[n] <= S，且不小于0。 - kv\_block\_table、score\_block\_table支持输入shape[B,ceil(Smax/block_size)]。Smax为每个Batch中最大的Sequence Length，即Smax=max(start\_pos)+S。 - cmp\_kv，输出shape为[B,ceil(S/cmp_ratio),D]：(compressed_tokens+pad0) + (compressed_tokens+pad1) + ... + (compressed_tokens+padN)。 - wkv\_proj，输出shape为[B,S,coff* D]。 - norm\_x，输出shape为[B,ceil(S/cmp_ratio),D]。 - norm\_rstd，输出shape为[B,ceil(S/cmp_ratio)]。 - 输入值域限制： - 该接口支持B、S泛化，且存在如下场景限制： - 部分长序列场景下，如果计算量过大可能会导致出现超过NPU内存的报错，注：这里计算量会受x输入shape的影响，值越大计算量越大。典型的长序列（即B、S的乘积或T较大）场景包括但不限于：
B S H
100 65525 4096
25 261120 4096
100 131072 4096
100 261120 4096
- 输入属性限制： - 支持D为128/512。 - 支持H为1K~10K，512对齐。 - 泛化支持block_size小于等于1024，16对齐。 - 支持cmp_ratio为4/128。支持如下三种情况： - C4A: D=512, coff=2, cmp_ratio=4； - C4Li: D=128, coff=2, cmp_ratio=4； - C128A: D=512, coff=1, cmp_ratio=128。 - 支持rotary_mode为2，Rope计算模式为interleave。 ## Atlas A3 推理系列产品 调用说明 - 单算子模式调用 ```python import torch import torch_npu import numpy as np import custom_ops import torch.nn as nn import math def get_seq_used_by_batch(batch_idx, S, seqused, cu_seqlens): if seqused is not None: return seqused[batch_idx] else: if cu_seqlens is not None: return cu_seqlens[batch_idx + 1] - cu_seqlens[batch_idx] else: return S data_type = torch.bfloat16 hidden_size = 4096 rope_head_dim = 64 norm_eps = 1e-6 coff = 1 # 1:no overlap 2:overlap cmp_ratio = 128 rotary_mode = 2 head_dim = 512 cu_seqlens = [0, 1] # ------------- B = 1 S = 1 S_max = 0 block_size = 128 start_pos = [8191] * B # (B,) start_p=8191 seqused = None # (B,), None时cu_seqlens的数据全部参与计算，否则按传参实际值计算 # BS是否合轴 bs_combine_flag = True update_flag = 1 if seqused is not None: seqused = torch.tensor(seqused).to(torch.int32) if start_pos is not None: start_pos = torch.tensor(start_pos).to(torch.int32) else: start_pos = torch.full((B,), start_p, dtype=torch.int32) if bs_combine_flag: if cu_seqlens is None: T = B * S if T !=0: cu_seqlens = torch.arange(0, T + 1, S, dtype=torch.int32) else: cu_seqlens = torch.zeros((B+1), dtype=torch.int32) else: cu_seqlens = torch.tensor(cu_seqlens).to(torch.int32) for i in range(B): if start_pos[i] + cu_seqlens[i + 1] - cu_seqlens[i] > S_max: S_max = start_pos[i] + cu_seqlens[i + 1] - cu_seqlens[i] else: cu_seqlens = None S_max = max(start_pos) + S ### ======================== gen input data start ============================= # page state max_block_num_per_batch = (S_max + block_size - 1) // block_size block_num = B * max_block_num_per_batch next_block_id = 1 print(f"max_block_num_per_batch: {max_block_num_per_batch}") block_table = torch.zeros(size=(B, max_block_num_per_batch), dtype=torch.int32) for i in range(B): # 需要读取state的范围 cur_start = start_pos[i] // cmp_ratio * cmp_ratio - cmp_ratio cur_end = start_pos[i] // cmp_ratio * cmp_ratio + cmp_ratio if start_pos[i] % cmp_ratio == 0: cur_end = start_pos[i] cur_end = min(cur_end, start_pos[i] + S) cur_start_block_id = (cur_start // block_size) if cur_start >= 0 else 0 cur_end_block_id = (cur_end - 1) // block_size for j in range(cur_start_block_id, cur_end_block_id + 1): block_table[i][j] = next_block_id next_block_id = next_block_id + 1 # 需要写入state的范围 end_pos = get_seq_used_by_batch(i, S, seqused, cu_seqlens) next_start = (start_pos[i] + end_pos) // cmp_ratio * cmp_ratio - cmp_ratio next_end = (start_pos[i] + end_pos) // cmp_ratio * cmp_ratio + cmp_ratio if (start_pos[i] + end_pos) % cmp_ratio == 0: next_end = start_pos[i] + end_pos next_end = min(next_end, start_pos[i] + end_pos) next_start_block_id = (next_start // block_size) if next_start >= 0 else 0 next_end_block_id = (next_end - 1) // block_size for j in range(next_start_block_id, next_end_block_id + 1): if block_table[i][j] == 0: block_table[i][j] = next_block_id next_block_id = next_block_id + 1 if B==0: kv_state = torch.tensor(np.random.uniform(-10, 10, (0, block_size, coff * head_dim))).to(torch.float32) score_state = torch.tensor(np.random.uniform(-10, 10, (0, block_size, coff * head_dim))).to(torch.float32) else: kv_state = torch.tensor(np.random.uniform(-10, 10, (torch.max(block_table) + 1, block_size, coff * head_dim))).to(torch.float32) score_state = torch.tensor(np.random.uniform(-10, 10, (torch.max(block_table) + 1, block_size, coff * head_dim))).to(torch.float32) # other input if bs_combine_flag: x_shape = (cu_seqlens[-1], hidden_size) rope_sin_shape = (min(x_shape[0], x_shape[0] // cmp_ratio + B), rope_head_dim) rope_cos_shape = rope_sin_shape else: x_shape = (B, S, hidden_size) rope_sin_shape = (B, (S + cmp_ratio - 1) // cmp_ratio, rope_head_dim) rope_cos_shape = rope_sin_shape x = torch.tensor(np.random.uniform(-10.0, 10.0, x_shape)).to(data_type).npu() wkv = torch.tensor(np.random.uniform(-10, 10, (coff * head_dim, hidden_size))).to(data_type).npu() wgate = torch.tensor(np.random.uniform(-10, 10, (coff * head_dim, hidden_size))).to(data_type).npu() ape = torch.tensor(np.random.uniform(-10, 10, (cmp_ratio, coff * head_dim))).to(torch.float32).npu() norm_weight = torch.tensor(np.random.uniform(-10, 10, (head_dim))).to(data_type).npu() rope_sin = torch.tensor(np.random.uniform(-1, 1, rope_sin_shape)).to(data_type).npu() rope_cos = torch.tensor(np.random.uniform(-1, 1, rope_cos_shape)).to(data_type).npu() kv_state = kv_state.npu() score_state = score_state.npu() block_table = block_table.npu() start_pos = torch.tensor(start_pos).to(torch.int32).npu() if cu_seqlens is not None: cu_seqlens = torch.tensor(cu_seqlens).to(torch.int32).npu() if seqused is not None: seqused = torch.tensor(seqused).to(torch.int32).npu() cmp_kv,_ ,_ ,_ ,_ = ( torch.ops.custom.compressor( x, wkv, wgate, kv_state, score_state, ape, norm_weight, rope_sin, rope_cos, kv_block_table = block_table, score_block_table = block_table, cu_seqlens = cu_seqlens, seqused = seqused, start_pos = start_pos, rope_head_dim = rope_head_dim, cmp_ratio = cmp_ratio, coff = coff, norm_eps = norm_eps, rotary_mode = rotary_mode ) ) ``` - aclgraph调用 ```python import torch import torch_npu import numpy as np import torch.nn as nn import torchair import custom_ops import math def get_seq_used_by_batch(batch_idx, S, seqused, cu_seqlens): if seqused is not None: return seqused[batch_idx] else: if cu_seqlens is not None: return cu_seqlens[batch_idx + 1] - cu_seqlens[batch_idx] else: return S data_type = torch.bfloat16 hidden_size = 4096 rope_head_dim = 64 norm_eps = 1e-6 coff = 1 # 1:no overlap 2:overlap cmp_ratio = 128 rotary_mode = 2 head_dim = 512 cu_seqlens = [0, 1] # ------------- B = 1 S = 1 S_max = 0 block_size = 128 start_pos = [8191] * B # (B,) start_p=8191 seqused = None # (B,), None时cu_seqlens的数据全部参与计算，否则按传参实际值计算 # BS是否合轴 bs_combine_flag = True update_flag = 1 if seqused is not None: seqused = torch.tensor(seqused).to(torch.int32) if start_pos is not None: start_pos = torch.tensor(start_pos).to(torch.int32) else: start_pos = torch.full((B,), start_p, dtype=torch.int32) if bs_combine_flag: if cu_seqlens is None: T = B * S if T !=0: cu_seqlens = torch.arange(0, T + 1, S, dtype=torch.int32) else: cu_seqlens = torch.zeros((B+1), dtype=torch.int32) else: cu_seqlens = torch.tensor(cu_seqlens).to(torch.int32) for i in range(B): if start_pos[i] + cu_seqlens[i + 1] - cu_seqlens[i] > S_max: S_max = start_pos[i] + cu_seqlens[i + 1] - cu_seqlens[i] else: cu_seqlens = None S_max = max(start_pos) + S ### ======================== gen input data start ============================= # page state max_block_num_per_batch = (S_max + block_size - 1) // block_size block_num = B * max_block_num_per_batch next_block_id = 1 print(f"max_block_num_per_batch: {max_block_num_per_batch}") block_table = torch.zeros(size=(B, max_block_num_per_batch), dtype=torch.int32) for i in range(B): # 需要读取state的范围 cur_start = start_pos[i] // cmp_ratio * cmp_ratio - cmp_ratio cur_end = start_pos[i] // cmp_ratio * cmp_ratio + cmp_ratio if start_pos[i] % cmp_ratio == 0: cur_end = start_pos[i] cur_end = min(cur_end, start_pos[i] + S) cur_start_block_id = (cur_start // block_size) if cur_start >= 0 else 0 cur_end_block_id = (cur_end - 1) // block_size for j in range(cur_start_block_id, cur_end_block_id + 1): block_table[i][j] = next_block_id next_block_id = next_block_id + 1 # 需要写入state的范围 end_pos = get_seq_used_by_batch(i, S, seqused, cu_seqlens) next_start = (start_pos[i] + end_pos) // cmp_ratio * cmp_ratio - cmp_ratio next_end = (start_pos[i] + end_pos) // cmp_ratio * cmp_ratio + cmp_ratio if (start_pos[i] + end_pos) % cmp_ratio == 0: next_end = start_pos[i] + end_pos next_end = min(next_end, start_pos[i] + end_pos) next_start_block_id = (next_start // block_size) if next_start >= 0 else 0 next_end_block_id = (next_end - 1) // block_size for j in range(next_start_block_id, next_end_block_id + 1): if block_table[i][j] == 0: block_table[i][j] = next_block_id next_block_id = next_block_id + 1 if B==0: kv_state = torch.tensor(np.random.uniform(-10, 10, (0, block_size, coff * head_dim))).to(torch.float32) score_state = torch.tensor(np.random.uniform(-10, 10, (0, block_size, coff * head_dim))).to(torch.float32) else: kv_state = torch.tensor(np.random.uniform(-10, 10, (torch.max(block_table) + 1, block_size, coff * head_dim))).to(torch.float32) score_state = torch.tensor(np.random.uniform(-10, 10, (torch.max(block_table) + 1, block_size, coff * head_dim))).to(torch.float32) # other input if bs_combine_flag: x_shape = (cu_seqlens[-1], hidden_size) rope_sin_shape = (min(x_shape[0], x_shape[0] // cmp_ratio + B), rope_head_dim) rope_cos_shape = rope_sin_shape else: x_shape = (B, S, hidden_size) rope_sin_shape = (B, (S + cmp_ratio - 1) // cmp_ratio, rope_head_dim) rope_cos_shape = rope_sin_shape x = torch.tensor(np.random.uniform(-10.0, 10.0, x_shape)).to(data_type).npu() wkv = torch.tensor(np.random.uniform(-10, 10, (coff * head_dim, hidden_size))).to(data_type).npu() wgate = torch.tensor(np.random.uniform(-10, 10, (coff * head_dim, hidden_size))).to(data_type).npu() ape = torch.tensor(np.random.uniform(-10, 10, (cmp_ratio, coff * head_dim))).to(torch.float32).npu() norm_weight = torch.tensor(np.random.uniform(-10, 10, (head_dim))).to(data_type).npu() rope_sin = torch.tensor(np.random.uniform(-1, 1, rope_sin_shape)).to(data_type).npu() rope_cos = torch.tensor(np.random.uniform(-1, 1, rope_cos_shape)).to(data_type).npu() kv_state = kv_state.npu() score_state = score_state.npu() block_table = block_table.npu() start_pos = torch.tensor(start_pos).to(torch.int32).npu() if cu_seqlens is not None: cu_seqlens = torch.tensor(cu_seqlens).to(torch.int32).npu() if seqused is not None: seqused = torch.tensor(seqused).to(torch.int32).npu() class CompressorNetwork(nn.Module): def __init__(self): super(CompressorNetwork, self).__init__() def forward(self, x, wkv, wgate, kv_state, score_state, ape, norm_weight, rope_sin, rope_cos, rope_head_dim, cmp_ratio, kv_block_table = None, score_block_table = None, cu_seqlens = None, seqused = None, start_pos = None, coff = 1, norm_eps = 1e-6, rotary_mode = 1): cmp_kv,_ ,_ ,_ ,_ = ( torch.ops.custom.compressor( x, wkv, wgate, kv_state, score_state, ape, norm_weight, rope_sin, rope_cos, kv_block_table = kv_block_table, score_block_table = score_block_table, cu_seqlens = cu_seqlens, seqused = seqused, start_pos = start_pos, rope_head_dim = rope_head_dim, cmp_ratio = cmp_ratio, coff = coff, norm_eps = norm_eps, rotary_mode = rotary_mode ) ) return cmp_kv from torchair.configs.compiler_config import CompilerConfig config = CompilerConfig() npu_backend = torchair.get_npu_backend(compiler_config=config) torch._dynamo.reset() npu_mode = torch.compile(CompressorNetwork(), fullgraph=True, backend=npu_backend, dynamic=False) cmp_kv = npu_mode( x, wkv, wgate, kv_state, score_state, ape, norm_weight, rope_sin, rope_cos, kv_block_table = block_table, score_block_table = block_table, cu_seqlens = cu_seqlens, seqused = seqused, start_pos = start_pos, rope_head_dim = rope_head_dim, cmp_ratio = cmp_ratio, coff = coff, norm_eps = norm_eps, rotary_mode = rotary_mode) ``` 更多使用示例见[pytest示例](./tests/pytest/README.md)。