[ascend]feat: support kv int8 (#2736)

* [ascend]feat: support kv int8 quant

* update code

* fix error of argument missing

* update params

* fix not iterable error when quant_meta is None.

* Update ascend_en_get_started.md for kvcache quant

* Update ascend_cn_get_started.md for kvcache quant

---------

Co-authored-by: jinminxi104 <[email protected]>

docs/en/get_started/ascend/get_started.md

@@ -136,3 +136,9 @@ lmdeploy lite auto_awq $HF_MODEL --work-dir $WORK_DIR --device npu
 ```
 
 Please check [supported_models](../../supported_models/supported_models.md) before use this feature.
+
+### int8 KV-cache Quantization
+
+Ascend backend has supported offline int8 KV-cache Quantization on eager mode.
+
+Please refer this [doc](https://github.com/DeepLink-org/dlinfer/blob/main/docs/quant/ascend_kv_quant.md) for details.

docs/zh_cn/get_started/ascend/get_started.md

@@ -133,3 +133,9 @@ lmdeploy lite auto_awq $HF_MODEL --work-dir $WORK_DIR --device npu
 ```
 
 支持的模型列表请参考[支持的模型](../../supported_models/supported_models.md)。
+
+### int8 KV-cache 量化
+
+昇腾后端现在支持了在eager模式下的离线int8 KV-cache量化。
+
+详细使用方式请请参考这篇[文章](https://github.com/DeepLink-org/dlinfer/blob/main/docs/quant/ascend_kv_quant.md)。

lmdeploy/messages.py

@@ -293,8 +293,11 @@ def __post_init__(self):
         assert self.device_type in [
             'cuda', 'ascend', 'maca'
         ], (f'invalid device_type: {self.device_type}')
-        if self.quant_policy > 0 and self.device_type != 'cuda':
-            assert False, 'kv cache quantization only works for CUDA.'
+        if self.quant_policy > 0 and self.device_type not in [
+                'cuda', 'ascend'
+        ]:
+            assert False, \
+                   'kv cache quantization only works for CUDA and ASCEND.'
 
 
 class ResponseType(enum.Enum):

lmdeploy/pytorch/backends/dlinfer/ascend/op_backend.py

@@ -1,5 +1,9 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from typing import Tuple
+import itertools
+import os
+import re
+from pathlib import Path
+from typing import Dict, Tuple
 
 import torch
 
@@ -11,6 +15,71 @@
 logger = get_logger('lmdeploy')
 
 
+class AscendKVQuantMeta:
+    has_set_value: bool = False
+    quant_meta: Dict = {}
+
+    @classmethod
+    def set_value(cls, device: str, dtype: torch.dtype, record_file: str,
+                  total_layers: int):
+        with open(record_file, 'r') as file:
+            data = file.read()
+        scale_offset_pairs = re.findall(
+            r'scale:\s*([\d\.\-]+)\s*offset:\s*(-?\d+)', data)
+        scale_offset_pairs = [(float(scale), float(offset))
+                              for scale, offset in scale_offset_pairs]
+        k_scales, v_scales, kv_scales = [], [], []
+        k_zeros, v_zeros, kv_zeros = [], [], []
+        if len(scale_offset_pairs) == total_layers:
+            for scale, offset in scale_offset_pairs:
+                k_scales.append(
+                    torch.tensor([scale], device=device, dtype=dtype))
+                v_scales.append(
+                    torch.tensor([scale], device=device, dtype=dtype))
+                kv_scales.append(
+                    torch.tensor([scale, scale], device=device, dtype=dtype))
+                k_zeros.append(
+                    torch.tensor([offset], device=device, dtype=dtype))
+                v_zeros.append(
+                    torch.tensor([offset], device=device, dtype=dtype))
+                kv_zeros.append(
+                    torch.tensor([offset, offset], device=device, dtype=dtype))
+        elif len(scale_offset_pairs) == total_layers * 2:
+            for i in range(total_layers):
+                scale_k, offset_k = scale_offset_pairs[2 * i]
+                scale_v, offset_v = scale_offset_pairs[2 * i + 1]
+                k_scales.append(
+                    torch.tensor([scale_k], device=device, dtype=dtype))
+                v_scales.append(
+                    torch.tensor([scale_v], device=device, dtype=dtype))
+                kv_scales.append(
+                    torch.tensor([scale_k, scale_v],
+                                 device=device,
+                                 dtype=dtype))
+                k_zeros.append(
+                    torch.tensor([offset_k], device=device, dtype=dtype))
+                v_zeros.append(
+                    torch.tensor([offset_v], device=device, dtype=dtype))
+                kv_zeros.append(
+                    torch.tensor([offset_k, offset_v],
+                                 device=device,
+                                 dtype=dtype))
+        else:
+            raise ValueError(
+                f'num of scale_offset_pairs({len(scale_offset_pairs)}) '
+                f'must match num of total_layers({total_layers})')
+
+        cls.quant_meta.update({
+            'k_scales': itertools.cycle(k_scales),
+            'k_zeros': itertools.cycle(k_zeros),
+            'v_scales': itertools.cycle(v_scales),
+            'v_zeros': itertools.cycle(v_zeros),
+            'kv_scales': itertools.cycle(kv_scales),
+            'kv_zeros': itertools.cycle(kv_zeros)
+        })
+        cls.has_set_value = True
+
+
 class AscendOpsBackend(DlinferOpsBackend):
     """ascend layer backend."""
     enable_graph = False
@@ -164,6 +233,21 @@ def get_total_slots():
                     .repeat_interleave(step_context.q_seqlens, 0)
             kv_seqlens = kv_seqlens_cpu
 
+        if not cls.enable_graph and step_context.kv_quant_policy == 8:
+            record_file = os.getenv('ASCEND_QUANT_RECORD_FILE')
+            assert record_file, 'please specify valid ASCEND_QUANT_RECORD_FILE'
+            path = Path(record_file)
+            is_path = path.is_absolute() or path.is_relative_to('/')
+            exists = path.exists()
+            if not (is_path and exists):
+                raise ValueError(
+                    'please specify valid ASCEND_QUANT_RECORD_FILE')
+            if not AscendKVQuantMeta.has_set_value:
+                total_layers = len(step_context.kv_caches)
+                AscendKVQuantMeta.set_value(step_context.block_offsets.device,
+                                            step_context.model_config.dtype,
+                                            record_file, total_layers)
+
         attn_meta_cls = cls.get_attention_metadata_cls()
         attn_metadata = attn_meta_cls(
             step_context.is_decoding,
@@ -177,6 +261,8 @@ def get_total_slots():
             is_unpaged_prefill=is_unpaged_prefill,
             max_q_seq_len=max_q_seq_len,
             max_kv_seq_len=max_kv_seq_len,
+            quant_policy=step_context.kv_quant_policy,
+            quant_meta=AscendKVQuantMeta.quant_meta,
         )
 
         step_context.attn_metadata = attn_metadata

lmdeploy/pytorch/backends/dlinfer/attention.py

@@ -1,6 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from dataclasses import dataclass
-from typing import Optional, Sequence
+from typing import Dict, Optional, Sequence
 
 from torch import Tensor
 
@@ -15,6 +15,7 @@ class DlinferAttentionMetadata(AttentionMetadata):
     is_unpaged_prefill: Optional[bool] = None
     max_q_seq_len: int = 1
     max_kv_seq_len: int = 1
+    quant_meta: Dict = None
 
 
 class DlinferAttentionImpl(AttentionImpl[DlinferAttentionMetadata]):
@@ -74,10 +75,37 @@ def forward(
         is_unpaged_prefill = attn_metadata.is_unpaged_prefill
         max_q_seq_len = attn_metadata.max_q_seq_len
         max_kv_seq_len = attn_metadata.max_kv_seq_len
+        quant_bits = attn_metadata.quant_policy
+        if attn_metadata.quant_meta is not None:
+            k_scales_zeros = [
+                next(attn_metadata.quant_meta['k_scales']),
+                next(attn_metadata.quant_meta['k_zeros'])
+            ] if 'k_scales' in attn_metadata.quant_meta else []
+            v_scales_zeros = [
+                next(attn_metadata.quant_meta['v_scales']),
+                next(attn_metadata.quant_meta['v_zeros'])
+            ] if 'v_scales' in attn_metadata.quant_meta else []
+            kv_scales = next(
+                attn_metadata.quant_meta['kv_scales']
+            ) if 'kv_scales' in attn_metadata.quant_meta else None
+            kv_zeros = next(
+                attn_metadata.quant_meta['kv_zeros']
+            ) if 'kv_zeros' in attn_metadata.quant_meta else None
+        else:
+            k_scales_zeros = []
+            v_scales_zeros = []
+            kv_scales = None
+            kv_zeros = None
 
         # fill kv cache
-        k_cache, v_cache = self.fill_kv_cache(key, value, k_cache, v_cache,
-                                              kv_start_indices)
+        k_cache, v_cache = self.fill_kv_cache(key,
+                                              value,
+                                              k_cache,
+                                              v_cache,
+                                              kv_start_indices,
+                                              k_scales_zeros=k_scales_zeros,
+                                              v_scales_zeros=v_scales_zeros,
+                                              quant_bits=quant_bits)
 
         if inplace:
             attn_output = query[..., :self.v_head_size]
@@ -103,6 +131,9 @@ def forward(
             block_size=block_size,
             attn_mask=attn_mask,
             is_unpaged_prefill=is_unpaged_prefill,
+            kv_scales=kv_scales,
+            kv_zeros=kv_zeros,
+            quant_bits=quant_bits,
         )
 
         return attn_output

lmdeploy/pytorch/config.py

@@ -77,6 +77,7 @@ class CacheConfig:
     max_prefill_token_num: int = 4096
     enable_prefix_caching: bool = False
     quant_policy: Literal[0, 4, 8] = 0
+    device_type: str = 'cuda'
 
     def __post_init__(self):
         """post init."""

lmdeploy/pytorch/engine/cache_engine.py

@@ -44,7 +44,13 @@ def __init__(
         self.num_layers = model_config.num_layers
         self.kv_cache_dtype = model_config.dtype
         if cache_config.quant_policy > 0:
-            self.kv_cache_dtype = torch.uint8
+            if self.cache_config.device_type in ['cuda']:
+                self.kv_cache_dtype = torch.uint8
+            elif self.cache_config.device_type in ['ascend', 'npu']:
+                self.kv_cache_dtype = torch.int8
+            else:
+                raise ValueError(
+                    f'unsupported device_type {self.cache_config.device_type}')
 
         # Initialize the cache.
         self.local_gpu_cache = self.allocate_gpu_cache()

lmdeploy/pytorch/engine/engine.py

@@ -130,6 +130,7 @@ def __init__(self,
             max_prefill_token_num=engine_config.max_prefill_token_num,
             enable_prefix_caching=engine_config.enable_prefix_caching,
             quant_policy=engine_config.quant_policy,
+            device_type=engine_config.device_type,
         )
 
         if not os.path.exists(model_path):

lmdeploy/pytorch/engine/model_agent.py

@@ -139,6 +139,7 @@ def model_forward(
         ctx_mgr = model.ctx_mgr
         context = ctx_mgr.build_context(
             inputs=inputs,
+            model_config=cache_engine.model_config,
             world_size=world_size,
             kv_caches=cache_engine.gpu_cache,
             kv_quant_policy=cache_engine.cache_config.quant_policy,

lmdeploy/pytorch/kernels/dlinfer/fill_kv_cache.py

@@ -1,4 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+from typing import Optional, Sequence
+
 import dlinfer.ops as ext_ops
 from torch import Tensor
 
@@ -9,7 +11,16 @@ def fill_kv_cache(
     key_caches: Tensor,
     value_caches: Tensor,
     kv_start_indices: Tensor,
+    k_scales_zeros: Sequence[Optional[Tensor]],
+    v_scales_zeros: Sequence[Optional[Tensor]],
+    quant_bits: int = 0,
 ):
     """fill key/value state to cache for paged attention."""
-    return ext_ops.fill_kv_cache(key_states, value_states, key_caches,
-                                 value_caches, kv_start_indices)
+    return ext_ops.fill_kv_cache(key_states,
+                                 value_states,
+                                 key_caches,
+                                 value_caches,
+                                 kv_start_indices,
+                                 k_scales_zeros=k_scales_zeros,
+                                 v_scales_zeros=v_scales_zeros,
+                                 quant_bits=quant_bits)

lmdeploy/pytorch/kernels/dlinfer/pagedattention.py

@@ -19,6 +19,9 @@ def prefill_attention(
     block_size: int,
     attn_mask: Sequence[Optional[Tensor]],
     is_unpaged_prefill: Optional[bool],
+    kv_scales: Optional[Tensor],
+    kv_zeros: Optional[Tensor],
+    quant_bits: Optional[int],
 ) -> Tensor:
     num_q_heads = query_states.shape[1]
     num_kv_heads = value_states.shape[1]
@@ -53,11 +56,25 @@ def prefill_attention(
             num_kv_heads,
             attn_mask,
             attn_output=attn_output,
+            kv_scales=kv_scales,
+            kv_zeros=kv_zeros,
+            quant_bits=quant_bits,
         )
 
 
-def paged_token_attention(q, k_cache, v_cache, attn_output, kv_seq_len,
-                          max_kv_seq_len, block_offsets, block_size):
+def paged_token_attention(
+    q,
+    k_cache,
+    v_cache,
+    attn_output,
+    kv_seq_len,
+    max_kv_seq_len,
+    block_offsets,
+    block_size,
+    kv_scales: Optional[Tensor],
+    kv_zeros: Optional[Tensor],
+    quant_bits: Optional[int],
+):
     num_q_heads, q_head_dim = q.shape[1:3]
     num_kv_heads = k_cache.shape[-1] // q_head_dim
     return ext_ops.paged_decode_attention(
@@ -71,6 +88,9 @@ def paged_token_attention(q, k_cache, v_cache, attn_output, kv_seq_len,
         num_q_heads,
         num_kv_heads,
         attn_output=attn_output,
+        kv_scales=kv_scales,
+        kv_zeros=kv_zeros,
+        quant_bits=quant_bits,
     )
 
 
@@ -91,6 +111,9 @@ def paged_attention_fwd(
     block_size: int,
     attn_mask: Sequence[Optional[Tensor]] = (),
     is_unpaged_prefill: Optional[bool] = None,
+    kv_scales: Optional[Tensor] = None,
+    kv_zeros: Optional[Tensor] = None,
+    quant_bits: Optional[int] = 0,
 ):
     if not is_decoding:
         return prefill_attention(
@@ -108,6 +131,9 @@ def paged_attention_fwd(
             block_size,
             attn_mask,
             is_unpaged_prefill,
+            kv_scales=kv_scales,
+            kv_zeros=kv_zeros,
+            quant_bits=quant_bits,
         )
     else:
         return paged_token_attention(
@@ -119,4 +145,7 @@ def paged_attention_fwd(
             max_kv_seq_len,
             block_offsets,
             block_size,
+            kv_scales=kv_scales,
+            kv_zeros=kv_zeros,
+            quant_bits=quant_bits,
         )