# SPDX-License-Identifier: Apache-2.0 """ This file implements common components for MLA implementations. First we define: Sq as Q sequence length Skv as KV sequence length MLA has two possible ways of computing, a data-movement friendly approach and a compute friendly approach, we generally want to use the compute friendly approach for "prefill" (i.e. the ratio Sq / Skv is "small", is near 1) and the data-movement friendly approach for "decode" (i.e. the ratio Sq / Skv is "large"). NOTE what we deem small and large is currently determined by if its labelled prefill or decode by the scheduler, but this is something we should probably tune. Main reference: DeepseekV2 paper, and FlashInfer Implementation (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551). Deepseek's MLA attention works the following way: * Use a single latent vector to represent the per-token entry of the KV cache. * For decode (i.e. the memory friendly approach) the attention "simulates" a multi-head attention, while the compute is similar to multi-query attention. Below is example of both paths assuming batchsize = 1 ## More Extent Definitions: C Context length, `Skv - Sq` H hidden size N number of attention heads Lq latent dimension for Q 1536 in DSV3 Lkv latent dimension for K/V 512 in DSV3 P nope dimension, no rope. 128 in DSV3 R rope dimension, goes through rope. 64 in DSV3 V V head dim. 128 in DSV3 ## Vector/Matrix Definitions h_t hidden states (input to attention) shape [Sq, H] q_c latent/compressed Q shape [Sq, Lq] q_nope uncompressed Q (no-rope) shape [Sq, N, P] q_pe uncompressed Q (rope) shape [Sq, N, R] kv_c latent/compressed KV shape [Skv, Lkv] k_pe decoupled k position embeddings shape [Skv, R] new_kv_c new kv_c from current iter shape [Sq, Lkv] new_k_pe new k_pe from current iter shape [Sq, R] cache_kv_c cached k_c from previous iters shape [C, Lkv] cache_k_pe cached k_pe from previous iters shape [C, R] W_DQ project h_t to q_c shape [H, Lq] W_UQ project q_c to q_nope shape [Lq, N * P] W_QR project q_c to q_pe shape [Lq, N * R] W_DKV project h_t to kv_c shape [H, Lkv] W_UK project kv_c to k_nope shape [Lkv, N, P] W_KR project h_t to k_pe shape [H, R] W_UV project kv_c to v shape [Lkv, N, V] W_O project v to h_t shape [N * V, H] ## Compute Friendly Approach (i.e. "_forward_prefill"): q_c = h_t @ W_DQ q_nope = (q_c @ W_UQ).view(Sq, N, P) q_pe = RoPE(q_c @ W_QR).view(Sq, N, R) new_kv_c = h_t @ W_DKV new_k_pe = RoPE(h_t @ W_KR) kv_c = torch.cat([new_kv_c, cache_kv_c], dim=0) k_pe = torch.cat([new_k_pe, cache_k_pe], dim=0) k_nope = (kv_c @ W_UK.view(Lkv, N * P)).view(Skv, N, P) v = (kv_c @ W_UV.view(Lkv, N * V)).view(Skv, N, V) // MHA with QK headdim = P + R // V headdim = V // spda_o shape [Sq, N, V] spda_o = scaled_dot_product_attention( torch.cat([q_nope, q_pe], dim=-1), torch.cat([k_nope, k_pe.unsqueeze(1).expand(-1, N, -1)], dim=-1), v ) return spda_o @ W_O NOTE: in the actual code, `kv_b_proj` is [W_UK; W_UV] concatnated per head `q_b_proj` is [W_UQ; W_QR] concatnated per head `out_proj` is W_O ## Data-Movement Friendly Approach (i.e. "_forward_decode"): Runtime q_c = h_t @ W_DQ q_nope = (q_c @ W_UQ).view(-1, N, P) ql_nope = einsum("snh,lnh->snl", q, W_UK) q_pe = RoPE(q_c @ W_QR).view(Sq, N, R) new_kv_c = h_t @ W_DKV new_k_pe = RoPE(h_t @ W_KR) kv_c = torch.cat([new_kv_c, cache_kv_c], dim=0) k_pe = torch.cat([new_k_pe, cache_k_pe], dim=0) // MQA with QK headdim = Lkv + R // V headdim = Lkv // spda_o shape [Sq, N, Lkv] // NOTE: this is less compute-friendly since Lkv > P // but is more data-movement friendly since its MQA vs MHA spda_o = scaled_dot_product_attention( torch.cat([ql_nope, q_pe], dim=-1), torch.cat([kv_c, k_pe], dim=-1), kv_c ) o = einsum("snl,lnv->snv", spda_o.reshape(-1, N, Lkv), W_UV) return o.view(-1, N * V) @ self.num_heads @ W_O ## Chunked Prefill For chunked prefill we want to use the compute friendly algorithm. We are assuming sufficiently large Sq / Skv ratio, in the future may want to switch to the data-movement friendly approach if the chunk (i.e. `Sq`) is small. However, the compute-friendly approach can potentially run out of memory if Skv is large due to: `k_nope = (kv_c @ W_UK).view(Skv, N, P)` To mitigate this, we chunk the computation of attention with respect to the current context (i.e. `cache_kv_c` and `cache_k_pe`) so that we can used a fixed workspace size. The chunked prefill approach is as follows: MCC Max chunk of context to process per iter, computed dynamically, used to bound the memory usage q_c = h_t @ W_DQ q_nope = (q_c @ W_UQ).view(Sq, N, P) q_pe = RoPE(q_c @ W_QR).view(Sq, N, R) new_kv_c = h_t @ W_DKV new_k_pe = RoPE(h_t @ W_KR) new_k_nope = (new_kv_c @ W_UK.view(Lkv, N * P)).view(Sq, N, P) new_v = (new_kv_c @ W_UV.view(Lkv, N * V)).view(Sq, N, V) // MHA between queries and new KV // with QK headdim = P + R // V headdim = V // curr_o shape [Sq, N, V] // curr_lse shape [N, Sq], this is just order FA returns curr_o, curr_lse = scaled_dot_product_attention( torch.cat([q_nope, q_pe], dim=-1), torch.cat([new_k_nope, new_k_pe.unsqueeze(1).expand(-1, N, -1)], dim=-1), new_v, casual=True, return_softmax_lse=True ) // Compute attention with the already existing context for chunk_idx in range(cdiv(C, MCC)): chunk_start = chunk_idx * MCC chunk_end = min(chunk_start + MCC, C) Sc = chunk_end - chunk_start cache_kv_c_chunk = cache_kv_c[chunk_start:chunk_end] cache_k_pe_chunk = cache_k_pe[chunk_start:chunk_end] cache_k_nope_chunk = (cache_kv_c_chunk @ W_UK).view(-1, N, P) cache_v_chunk = (cache_kv_c_chunk @ W_UV).view(-1, N, V) chunk_o, chunk_lse = scaled_dot_product_attention( torch.cat([q_nope, q_pe], dim=-1), torch.cat([cache_k_nope_chunk, cache_k_pe_chunk.unsqueeze(1).expand(-1, N, -1)], dim=-1), cache_v_chunk, casual=False, return_softmax_lse=True ) curr_o, curr_lse = merge_attn_states( suffix_output=curr_o, suffix_lse=curr_lse, prefix_output=chunk_o, prefix_lse=chunk_lse, ) return curr_o @ W_O """ import functools from abc import abstractmethod from dataclasses import dataclass from typing import TYPE_CHECKING, Any, Generic, Optional, TypeVar import torch from vllm import _custom_ops as ops import vllm.envs as envs from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer, AttentionMetadata, MLAAttentionImpl) from vllm.attention.ops.triton_merge_attn_states import merge_attn_states from vllm.logger import init_logger from vllm.model_executor.layers.linear import (ColumnParallelLinear, LinearBase, RowParallelLinear, UnquantizedLinearMethod) from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding from vllm.platforms import current_platform from vllm.utils import cdiv, round_down from vllm.vllm_flash_attn.fa_utils import get_flash_attn_version # try: # from vllm.vllm_flash_attn import flash_attn_varlen_func # except ImportError: # # For rocm use upstream flash attention # from flash_attn import flash_attn_varlen_func from ixformer.contrib.vllm_flash_attn import flash_attn_varlen_func, merge_attn_states if TYPE_CHECKING: from vllm.v1.core.sched.output import SchedulerOutput from vllm.v1.worker.gpu_input_batch import InputBatch from vllm.v1.worker.gpu_model_runner import GPUModelRunner from vllm.model_executor.layers.quantization.utils.quant_utils import ( scaled_quantize, scaled_dequantize) import ixformer.inference.functions as ixf_ops logger = init_logger(__name__) class MLACommonBackend(AttentionBackend): accept_output_buffer: bool = True @staticmethod def get_name() -> str: return "TRITON_MLA_VLLM_V1" @staticmethod def get_metadata_cls() -> type["AttentionMetadata"]: return MLACommonMetadata @staticmethod def get_builder_cls() -> type["MLACommonMetadataBuilder"]: return MLACommonMetadataBuilder @staticmethod def get_kv_cache_shape( num_blocks: int, block_size: int, num_kv_heads: int, # assumed to be 1 for MLA head_size: int, ) -> tuple[int, ...]: return (num_blocks, block_size, head_size) @staticmethod def get_supported_head_sizes() -> list[int]: return [576] @staticmethod def use_cascade_attention(*args, **kwargs) -> bool: return False @dataclass class MLACommonPrefillMetadata: """ Prefill Specific Metadata """ @dataclass class ChunkedContextMetadata: # New for MLA (compared to FlashAttention) # For handling chunked prefill cu_seq_lens: torch.Tensor starts: torch.Tensor seq_tot: list[int] max_seq_lens: list[int] workspace: torch.Tensor # Input positions for rotrary embeddings since for MLA the rotary # position embeddings are applied inside the attention backend input_positions: torch.Tensor block_table: torch.Tensor query_start_loc: torch.Tensor max_query_len: int chunked_context: Optional[ChunkedContextMetadata] = None @dataclass class MLACommonDecodeMetadata: # Input positions for rotrary embeddings since for MLA the rotary # position embeddings are applied inside the attention backend input_positions: torch.Tensor block_table: torch.Tensor seq_lens: torch.Tensor max_decode_seq_len: int use_cuda_graph: bool D = TypeVar("D", bound=MLACommonDecodeMetadata) @dataclass class MLACommonMetadata(Generic[D]): """Metadata for MLACommon. NOTE: Please read the comment at the top of the file before trying to understand this class """ # NOTE(sang): Definition of context_len, query_len, and seq_len. # |---------- N-1 iteration --------| # |---------------- N iteration ---------------------| # |- tokenA -|......................|-- newTokens ---| # |---------- context_len ----------| # |-------------------- seq_len ---------------------| # |-- query_len ---| num_actual_tokens: int # Number of tokens excluding padding. query_start_loc: torch.Tensor slot_mapping: torch.Tensor # New for MLA (compared to FlashAttention) # For handling prefill decode split num_decodes: int num_decode_tokens: int num_prefills: int # For logging. num_input_tokens: int = 0 # Number of tokens including padding. # The dimension of the attention heads head_dim: Optional[int] = None decode: Optional[D] = None prefill: Optional[MLACommonPrefillMetadata] = None def __post_init__(self): supported_head_sizes = MLACommonBackend.get_supported_head_sizes() if self.head_dim is not None and self.head_dim \ not in supported_head_sizes: raise ValueError( f"Only {supported_head_sizes} are supported for head_dim,", f"received {self.head_dim}.") M = TypeVar("M", bound=MLACommonMetadata) class MLACommonMetadataBuilder(Generic[M]): """ NOTE: Please read the comment at the top of the file before trying to understand this class """ def __init__(self, runner: "GPUModelRunner", metadata_cls: Optional[type[M]] = None): self.metadata_cls = metadata_cls \ if metadata_cls is not None else MLACommonMetadata self.runner = runner scheduler_config = runner.scheduler_config model_config = runner.model_config cache_config = runner.cache_config self.chunked_prefill_enabled = scheduler_config.chunked_prefill_enabled if self.chunked_prefill_enabled: self.chunked_prefill_workspace_size = min( # Max sure there is enough for 8 full length request or at least # 4 pages of cache per request max( 8 * model_config.max_model_len, 4 * scheduler_config.max_num_seqs * cache_config.block_size), # For long-context models try not to over-allocate limiting # kv-cache space, limiting it to 64k tokens, # which would result in the workspace being: # 2*(576)*(64*1024) = 144mb # (assuming 576 MLA head dim, and fp16) # which would result in up-projected context being # 2*(192*128)*(64*1024) = 3gb # (assuming 192 QK head dim, 128 heads, and fp16) 128 * 1024) assert self.chunked_prefill_workspace_size >= \ scheduler_config.max_num_seqs * cache_config.block_size self.chunked_prefill_workspace = torch.empty( (self.chunked_prefill_workspace_size, model_config.get_head_size()), dtype=model_config.dtype, device=runner.device, ) self.page_size = self.runner.block_size def reorder_batch(self, input_batch: "InputBatch", scheduler_output: "SchedulerOutput") -> bool: # We now want to reorder the batch so that the "decode" requests are and # the front and the "prefill" requests are at the using the least amount # swaps possible. (NOTE for now we loosely use "decode" to mean requests # where attention is likely memory-bound and "prefill" to mean requests # where attention is likely compute-bound, TODO(lucas): figure out a # better naming here) decodes = [] prefills = [] num_decode_tokens = 0 num_prefill_tokens = 0 for i, req_id in enumerate(input_batch.req_ids): num_tokens = scheduler_output.num_scheduled_tokens[req_id] # for now treat 1 scheduled token as "decode" even if its not, # we should update this to something like < 8 in the future but # currently the TritonMLA._forward_decode only supports # num_tokens = 1 if num_tokens == 1: decodes.append(i) num_decode_tokens += num_tokens else: prefills.append(i) num_prefill_tokens += num_tokens # We hope that this is fairly minimal since decodes # should be around for a number of iterations so hopefully they are # relatively stationary (and new request are generally appended to the # persistent batch so already should be at the back) # To achieve this we loop over the decodes in descending order and # the prefills in ascending order. We swap decodes from the "back" # i.e. past where the last decode should be in the reodorered with # prefills from the front of the batch. # `decodes` and `prefills` are already in ascending order just based on # the above loop num_decodes = len(decodes) num_prefills = len(prefills) first_prefill = 0 modified_batch = False for i in range(1, min(num_decodes, num_prefills) + 1): # If the decode is at the "back" of the batch, i, we can swap it # with the prefill closest to the front of the batch if decodes[num_decodes - i] >= num_decodes: input_batch.swap_states(prefills[first_prefill], decodes[num_decodes - i]) first_prefill += 1 modified_batch = True else: break # Save for next `build` call # TODO(lucas): this is a bit of a hack, we should probably have a # better way of doing this self._num_decodes = num_decodes self._num_prefills = num_prefills self._num_decode_tokens = num_decode_tokens self._num_prefill_tokens = num_prefill_tokens return modified_batch def _build_decode(self, input_positions: torch.Tensor, block_table: torch.Tensor, seq_lens: torch.Tensor, max_decode_seq_len: int, use_cuda_graph: bool): return MLACommonDecodeMetadata( input_positions=input_positions, block_table=block_table, seq_lens=seq_lens, max_decode_seq_len=max_decode_seq_len, use_cuda_graph=use_cuda_graph, ) def build(self, num_reqs: int, num_actual_tokens: int, max_query_len: int, common_prefix_len: int) -> M: assert self._num_decodes + self._num_prefills == num_reqs # Note(simon): be careful about the CPU <> GPU memory movement in this # function. We should avoid GPU -> CPU sync as much as possible because # it blocks on all previous kernels. device = self.runner.device block_table = ( self.runner.input_batch.block_table.get_device_tensor()[:num_reqs]) query_start_loc = self.runner.query_start_loc_cpu[:num_reqs + 1].to( device, non_blocking=True) slot_mapping = self.runner.slot_mapping_cpu[:num_actual_tokens].to( device, non_blocking=True).long() input_positions = self.runner.positions_cpu[:num_actual_tokens].to( device, non_blocking=True).long() seq_lens_cpu = self.runner.seq_lens_cpu[:num_reqs] seq_lens = seq_lens_cpu.to(device, non_blocking=True) max_query_len = seq_lens_cpu.max().item() prefill_metadata = None if self._num_prefills > 0: reqs_start = self._num_decodes # prefill_start tokens_start = self._num_decode_tokens context_lens_cpu = self.runner.input_batch.\ num_computed_tokens_cpu_tensor[reqs_start:num_reqs] max_context_len_cpu = context_lens_cpu.max().item() num_prefills_with_context_cpu = (context_lens_cpu > 0).sum().item() chunked_context_metadata = None if self.chunked_prefill_enabled and self._num_prefills > 0 \ and max_context_len_cpu > 0: # NOTE: it is recommend you read the `Chunked Prefill` section # in the comment at the top of the file before trying to # understand the following code # currently we allocate an equal amount of workspace for each # prefill in the batch, we could probably use a more advanced # algorithm here and allocate more workspace to prefills with # longer context lengths max_context_chunk = (self.chunked_prefill_workspace_size // num_prefills_with_context_cpu) # align max_context_chunk to page_size by rounding down, # currently the `gather_cache` kernel cannot handle # `context_chunk_starts` that are not aligned to page_size max_context_chunk = round_down(max_context_chunk, self.page_size) assert max_context_chunk > 0 num_chunks = cdiv(max_context_len_cpu, max_context_chunk) # if `max_context_chunk = 256`, `num_chunks = 3`, and # `num_prefills_with_context = 4`, create a tensor that looks # like # [[0, 0, 0, 0], [256, 256, 256, 256], [512, 512, 512, 512]] # Note(simon): this is done in CPU because of downstream's # of `to_list`. chunk_starts = \ torch.arange(num_chunks, dtype=torch.int32) \ .unsqueeze(1).expand(-1, self._num_prefills) \ * max_context_chunk chunk_ends = torch.min(context_lens_cpu.unsqueeze(0), chunk_starts + max_context_chunk) chunk_seq_lens = (chunk_ends - chunk_starts).clamp(min=0) cu_seq_lens_cpu = torch.zeros(num_chunks, self._num_prefills + 1, dtype=torch.int32, pin_memory=True) torch.cumsum(chunk_seq_lens, dim=1, out=cu_seq_lens_cpu[:, 1:], dtype=torch.int32) chunked_context_metadata = \ MLACommonPrefillMetadata.ChunkedContextMetadata( cu_seq_lens=cu_seq_lens_cpu.to(device, non_blocking=True), starts=chunk_starts.to(device, non_blocking=True), seq_tot=chunk_seq_lens.sum(dim=1).tolist(), max_seq_lens=chunk_seq_lens.max(dim=1).values.tolist(), workspace=self.chunked_prefill_workspace, ) assert max(chunked_context_metadata.max_seq_lens) <= \ self.chunked_prefill_workspace_size prefill_metadata = MLACommonPrefillMetadata( input_positions=input_positions[tokens_start:], block_table=block_table[reqs_start:, ...], query_start_loc=query_start_loc[reqs_start:] - query_start_loc[reqs_start], max_query_len=max_query_len, chunked_context=chunked_context_metadata, ) decode_metadata = None if self._num_decodes > 0: decode_metadata = self._build_decode( input_positions=input_positions[:self._num_decode_tokens], block_table=block_table[:self._num_decodes, ...], seq_lens=seq_lens[:self._num_decodes], max_decode_seq_len=max_query_len, use_cuda_graph=False, ) return self.metadata_cls( num_actual_tokens=num_actual_tokens, query_start_loc=query_start_loc, slot_mapping=slot_mapping, head_dim=self.runner.model_config.get_head_size(), # MLACommonMetadata Chunk prefill specific num_decodes=self._num_decodes, num_decode_tokens=self._num_decode_tokens, num_prefills=self._num_prefills, prefill=prefill_metadata, decode=decode_metadata, ) class MLACommonImpl(MLAAttentionImpl[M], Generic[M]): """ NOTE: Please read the comment at the top of the file before trying to understand this class """ def __init__( self, num_heads: int, head_size: int, scale: float, num_kv_heads: int, alibi_slopes: Optional[list[float]], sliding_window: Optional[int], kv_cache_dtype: str, blocksparse_params: Optional[dict[str, Any]], logits_soft_cap: Optional[float], attn_type: str, # MLA Specific Arguments q_lora_rank: Optional[int], kv_lora_rank: int, qk_nope_head_dim: int, qk_rope_head_dim: int, qk_head_dim: int, v_head_dim: int, rotary_emb: RotaryEmbedding, # q_proj should be q_b_proj if q_lora_rank is not None, but from an # attention backend perspective we rely on the layer to pass in the # correct matrix q_proj: ColumnParallelLinear, kv_b_proj: ColumnParallelLinear, o_proj: RowParallelLinear, ) -> None: self.num_heads = num_heads self.head_size = head_size self.scale = float(scale) self.num_kv_heads = num_kv_heads self.kv_cache_dtype = kv_cache_dtype self.q_lora_rank = q_lora_rank self.kv_lora_rank = kv_lora_rank self.qk_nope_head_dim = qk_nope_head_dim self.qk_rope_head_dim = qk_rope_head_dim self.qk_head_dim = qk_head_dim self.v_head_dim = v_head_dim # Hack for V1 for now to avoid torch library overhead (since we are # already inside an attention custom op), pull out the forward # method from the rotary embedding and call it directly # TODO(lucas): we should probably find a cleaner way to do this self.rotary_emb = rotary_emb # self.rotary_emb = rotary_emb.forward_native # if current_platform.is_cuda(): # self.rotary_emb = rotary_emb.forward_cuda self.q_proj = q_proj self.kv_b_proj = kv_b_proj self.o_proj = o_proj self.vllm_flash_attn_version = get_flash_attn_version() # Handle the differences between the flash_attn_varlen from flash_attn # and the one from vllm_flash_attn. The former is used on RoCM and the # latter has an additional parameter to control FA2 vs FA3 self.flash_attn_varlen_func = flash_attn_varlen_func if self.vllm_flash_attn_version is not None: self.flash_attn_varlen_func = \ functools.partial(flash_attn_varlen_func, fa_version=self.vllm_flash_attn_version) def _v_up_proj_and_o_proj(self, x): # # Convert from (B, N, L) to (N, B, L) # x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1) # # Multiply (N, B, L) x (N, L, V) -> (N, B, V) # x = torch.bmm(x, self.W_UV) # # Convert from (N, B, V) to (B, N * V) # x = x.transpose(0, 1).reshape(-1, self.num_heads * self.v_head_dim) # return self.o_proj(x)[0] x = torch.einsum("bnl,lnv->bnv", x, self.W_UV) return self.o_proj(x.reshape(-1, self.num_heads * self.v_head_dim))[0] # Return `ql_nope`, `q_pe` def _q_proj_and_k_up_proj(self, x): q_nope, q_pe = self.q_proj(x)[0]\ .view(-1, self.num_heads, self.qk_head_dim)\ .split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) # # Convert from (B, N, P) to (N, B, P) # q_nope = q_nope.transpose(0, 1) # # Multiply (N, B, P) x (N, P, L) -> (N, B, L) # ql_nope = torch.bmm(q_nope, self.W_UK_T) # # Convert from (N, B, L) to (B, N, L) # return ql_nope.transpose(0, 1), q_pe return torch.einsum("bnp,lnp->bnl", q_nope, self.W_UK).view(-1, self.num_heads, self.kv_lora_rank), q_pe def process_weights_after_loading(self, act_dtype: torch.dtype): def get_layer_weight(layer): WEIGHT_NAMES = ("weight", "qweight", "weight_packed") for attr in WEIGHT_NAMES: if hasattr(layer, attr): return getattr(layer, attr) raise AttributeError( f"Layer '{layer}' has no recognized weight attribute:" f" {WEIGHT_NAMES}.") def get_and_maybe_dequant_weights(layer: LinearBase): if not isinstance(layer.quant_method, UnquantizedLinearMethod): # we already prepare a dequant weight for GGUF, skip it. if layer.quant_method.__class__.__name__ == "GGUFLinearMethod": weight = layer.weight.clone() del layer.weight return weight if layer.quant_method.__class__.__name__ == "AWQMarlinLinearMethod": from ixformer.inference.functions import ref_wui4a16 return ref_wui4a16(None, layer.qweight, layer.scales, layer.qzeros, None, layer.quant_method.quant_config.group_size, only_return_weight=True) # for W8A8, we directly dequantize it here to avoiding quantization errors if hasattr(layer, "scheme") and layer.scheme.__class__.__name__ == "CompressedTensorsW8A8Int8" and not layer.scheme.is_static_input_scheme: quant_weight = layer.weight.T # output, input scales = layer.weight_scale return scaled_dequantize(quant_weight, scales, (1, -1), act_dtype) # NOTE: This should only be used offline, since it's O(N^3) eye = torch.eye(layer.input_size_per_partition, dtype=act_dtype, device=get_layer_weight(layer).device) dequant_weights = layer.quant_method.apply(layer, eye, bias=None) del eye # standardize to (output, input) return dequant_weights.T return layer.weight back_to_vllm = False weight_dtype = get_layer_weight(self.kv_b_proj).dtype assert get_layer_weight(self.q_proj).dtype == weight_dtype # when use customize forward, we do not care the specific data types and shape, just reset the # _v_up_proj_and_o_proj and _q_proj_and_k_up_proj funs to get the correct results. if envs.VLLM_MLA_CUSTOMIZE: layer = self.kv_b_proj quant_method = layer.quant_method.__class__.__name__ if hasattr(layer, "scheme") and layer.scheme.__class__.__name__ == "CompressedTensorsW8A8Int8": # must be as same as kv_b_proj assert hasattr(self.q_proj, "scheme") and self.q_proj.scheme.__class__.__name__ == "CompressedTensorsW8A8Int8" assert hasattr(self.o_proj, "scheme") and self.o_proj.scheme.__class__.__name__ == "CompressedTensorsW8A8Int8" kv_b_proj_weight = self.kv_b_proj.weight # [i，o] kv_b_proj_weight_scale = self.kv_b_proj.weight_scale # [o, 1] kv_b_proj_weight = kv_b_proj_weight.view( self.kv_lora_rank, self.num_heads, self.qk_nope_head_dim + self.v_head_dim, ) W_UK, W_UV = kv_b_proj_weight.split( [self.qk_nope_head_dim, self.v_head_dim], dim=-1) kv_b_proj_weight_scale = kv_b_proj_weight_scale.view( self.num_heads, self.qk_nope_head_dim + self.v_head_dim, 1, ) W_UK_S, W_UV_S = kv_b_proj_weight_scale.split( [self.qk_nope_head_dim, self.v_head_dim], dim=1) # self.W_UK = W_UK # [kv_lora_rank, n, qk_nope] self.W_UK = ixf_ops.marlin_w8_weight_repack( weights=W_UK.permute(1, 2, 0).contiguous(), weight_format="int8", reformat="k16n16_grouped_n", ) self.W_UK_S = W_UK_S.contiguous() # [n, qk_nope, 1] # self.W_UV = W_UV # [kv_lora_rank, n, v_head_dim] self.W_UV = ixf_ops.marlin_w8_weight_repack( weights=W_UV.permute(1, 0, 2).contiguous(), weight_format="int8", reformat="k16n16", ) self.W_UV_S = W_UV_S.contiguous() # [n, v_head_dim, 1] def _v_up_proj_and_o_proj_w8a8(self, x): # x: b, n, kv_lora # W_UV = (self.W_UV * self.W_UV_S.view(1, self.num_heads, self.v_head_dim)).to(x.dtype) # x = torch.einsum("bnl,lnv->bnv", x, W_UV) res = ixf_ops.marlin_w8a16( x, self.W_UV, self.W_UV_S, group_size=-1, format="k16n16", batch_first=False, ) return self.o_proj(res.reshape(-1, self.num_heads * self.v_head_dim))[0] def _q_proj_and_k_up_proj_w8a8(self, x): # x: b, q_lora x = self.q_proj(x)[0].view(-1, self.num_heads, self.qk_head_dim) x_nope, x_pe = x.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) # W_UK = (self.W_UK * self.W_UK_S.view(1, self.num_heads, self.qk_nope_head_dim)).to(x.dtype) # return torch.einsum("bnp,lnp->bnl", x_nope, W_UK).view(-1, self.num_heads, self.kv_lora_rank), x_pe res = ixf_ops.marlin_w8a16( x_nope, self.W_UK, self.W_UK_S, group_size=-1, format="k16n16_grouped_n", batch_first=False, ) return res, x_pe self._v_up_proj_and_o_proj = functools.partial(_v_up_proj_and_o_proj_w8a8, self) self._q_proj_and_k_up_proj = functools.partial(_q_proj_and_k_up_proj_w8a8, self) elif quant_method == "AWQMarlinLinearMethod": # ===== W_UK & W_UV use W4A16 ===== def split_4bit_matrix(qweight, scales, qzeros, num_heads, head_dim_1, head_dim_2): assert head_dim_1 % 8 == 0 assert head_dim_2 % 8 == 0 qweight = qweight.view(-1, num_heads, (head_dim_1 + head_dim_2)//8) w1, w2 = qweight.split([head_dim_1//8, head_dim_2//8], dim=-1) scales = scales.view(-1, num_heads, head_dim_1 + head_dim_2) s1, s2 = scales.split([head_dim_1, head_dim_2], dim=-1) qzeros = qzeros.view(-1, num_heads, (head_dim_1 + head_dim_2)//8) z1, z2 = qzeros.split([head_dim_1//8, head_dim_2//8], dim=-1) return (w1.contiguous().view(w1.shape[0], -1), s1.contiguous().view(s1.shape[0], -1), z1.contiguous().view(z1.shape[0], -1), w2.contiguous().view(w2.shape[0], -1), s2.contiguous().view(s2.shape[0], -1), z2.contiguous().view(z2.shape[0], -1)) # split W_UK and W_UV ( W_UK, W_UK_S, W_UK_Z, W_UV, W_UV_S, W_UV_Z ) = split_4bit_matrix( self.kv_b_proj.qweight, self.kv_b_proj.scales, self.kv_b_proj.qzeros, self.num_heads, self.qk_nope_head_dim, self.v_head_dim ) # repack W_UK W_UK = W_UK.reshape(self.kv_lora_rank, self.num_heads, -1).permute(1, 2, 0).contiguous() W_UK_S = W_UK_S.reshape(W_UK_S.shape[0], self.num_heads, -1).permute(1, 0, 2).contiguous() W_UK_Z = W_UK_Z.reshape(W_UK_Z.shape[0], self.num_heads, -1).permute(1, 0, 2).contiguous() ( self.W_UK, self.W_UK_S, self.W_UK_Z ) = ixf_ops.marlin_w4_weight_repack( weights = W_UK, scales = W_UK_S, zeros = W_UK_Z, weight_format = "gptq_grouped_n", reformat = "k16n32_grouped_n", ) # repack W_UV W_UV = W_UV.reshape(self.kv_lora_rank, self.num_heads, -1).permute(1, 0, 2).contiguous() W_UV_S = W_UV_S.reshape(W_UV_S.shape[0], self.num_heads, -1).permute(1, 0, 2).contiguous() W_UV_Z = W_UV_Z.reshape(W_UV_Z.shape[0], self.num_heads, -1).permute(1, 0, 2).contiguous() ( self.W_UV, self.W_UV_S, self.W_UV_Z ) = ixf_ops.marlin_w4_weight_repack( weights = W_UV, scales = W_UV_S, zeros = W_UV_Z, weight_format = "awq", reformat = "k16n32", ) self.w4_group_size = self.q_proj.quant_method.quant_config.group_size # ===== W_UK & W_UV use W16A16 ===== # kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T # kv_b_proj_weight = kv_b_proj_weight.view( # self.kv_lora_rank, # self.num_heads, # self.qk_nope_head_dim + self.v_head_dim, # ) # self.W_UK_fp16, self.W_UV_fp16 = kv_b_proj_weight.split( # [self.qk_nope_head_dim, self.v_head_dim], dim=-1) def _v_up_proj_and_o_proj_w4a16(self, x): # W16A16 # res = torch.einsum("bnl,lnv->bnv", x, self.W_UV_fp16) # W4A16 res = ixf_ops.marlin_w4a16( x, self.W_UV, self.W_UV_S, self.W_UV_Z, group_size=self.w4_group_size, format="k16n32", batch_first=False) return self.o_proj(res.view(-1, self.num_heads * self.v_head_dim))[0] def _q_proj_and_k_up_proj_w4a16(self, x): x = self.q_proj(x)[0].view(-1, self.num_heads, self.qk_head_dim) x_nope, x_pe = x.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) # W16A16 # return torch.einsum("bnp,lnp->bnl", x_nope, self.W_UK_fp16).view(-1, self.num_heads, self.kv_lora_rank), x_pe # W4A16 res = ixf_ops.marlin_w4a16( x_nope, self.W_UK, self.W_UK_S, self.W_UK_Z, group_size=self.w4_group_size, format="k16n32_grouped_n", batch_first=False, ) return res, x_pe self._v_up_proj_and_o_proj = functools.partial(_v_up_proj_and_o_proj_w4a16, self) self._q_proj_and_k_up_proj = functools.partial(_q_proj_and_k_up_proj_w4a16, self) else: print(f"Custom MLA for quant method: {layer.quant_method.__class__.__name__} is not supported, will use the vllm official impl.") back_to_vllm = True else: back_to_vllm = True if back_to_vllm: # we currently do not have quantized bmm's which are needed for # `W_UV` and `W_UK_T`, we we just store fp16/bf16 copies and perform # the bmm's in 16-bit, the extra memory overhead of this is fairly low kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T assert kv_b_proj_weight.shape == ( self.kv_lora_rank, self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), ( f"{kv_b_proj_weight.shape=}, " f"{self.kv_lora_rank=}, " f"{self.num_heads=}, " f"{self.qk_nope_head_dim=}, " f"{self.v_head_dim=}") kv_b_proj_weight = kv_b_proj_weight.view( self.kv_lora_rank, self.num_heads, self.qk_nope_head_dim + self.v_head_dim, ) W_UK, W_UV = kv_b_proj_weight.split( [self.qk_nope_head_dim, self.v_head_dim], dim=-1) # # Convert from (L, N, V) to (N, L, V) # self.W_UV = W_UV.transpose(0, 1) # # Convert from (L, N, P) to (N, P, L) # self.W_UK_T = W_UK.permute(1, 2, 0) self.W_UV = W_UV self.W_UK = W_UK def _compute_prefill_context( self, q: torch.Tensor, kv_c_and_k_pe_cache: torch.Tensor, kv_c_and_k_pe_cache_scale: torch.Tensor, attn_metadata: MLACommonMetadata, ): assert attn_metadata.prefill is not None prefill_metadata = attn_metadata.prefill assert prefill_metadata.chunked_context is not None output = None iters = len(prefill_metadata.chunked_context.seq_tot) workspace = prefill_metadata.chunked_context.workspace for i in range(iters): toks = prefill_metadata.chunked_context.seq_tot[i] if envs.VLLM_USE_INT8_MLA: ops.gather_cache_int8( src_cache=kv_c_and_k_pe_cache, src_cache_scale=kv_c_and_k_pe_cache_scale, kv_lora_rank = self.kv_lora_rank, dst=workspace, block_table=prefill_metadata.block_table, cu_seq_lens=prefill_metadata.chunked_context.cu_seq_lens[i], batch_size=attn_metadata.num_prefills, seq_starts=prefill_metadata.chunked_context.starts[i], ) else: ops.gather_cache( src_cache=kv_c_and_k_pe_cache, dst=workspace, block_table=prefill_metadata.block_table, cu_seq_lens=prefill_metadata.chunked_context.cu_seq_lens[i], batch_size=attn_metadata.num_prefills, seq_starts=prefill_metadata.chunked_context.starts[i], ) kv_c_normed = workspace[:toks]\ [..., :self.kv_lora_rank].contiguous() k_pe = workspace[:toks]\ [..., self.kv_lora_rank:].unsqueeze(1) kv_nope = self.kv_b_proj(kv_c_normed)[0].view( \ -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim) k_nope, v = kv_nope\ .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))), dim=-1) attn_output = torch.empty(q.shape[0], self.num_heads, self.v_head_dim, dtype=q.dtype, device=q.device) # For MLA the v head dim is smaller than qk head dim so we pad # out v with 0s to match the qk head dim # v_padded = torch.nn.functional.pad(v, # [0, q.shape[-1] - v.shape[-1]], # value=0) attn_output, attn_softmax_lse = self.flash_attn_varlen_func( q=q, k=k, v=v, cu_seqlens_q=prefill_metadata.query_start_loc, cu_seqlens_k=prefill_metadata.chunked_context.cu_seq_lens[i], max_seqlen_q=prefill_metadata.max_query_len, max_seqlen_k=prefill_metadata.chunked_context.max_seq_lens[i], softmax_scale=self.scale, causal=False, # Context is unmasked return_softmax_lse=True, out=attn_output, ) if output is None: output = attn_output output_lse = attn_softmax_lse else: output_tmp = torch.empty_like(output) output_lse_tmp = torch.empty_like(output_lse) merge_attn_states( output=output_tmp, output_lse=output_lse_tmp, prefix_output=output, prefix_lse=output_lse, suffix_output=attn_output, suffix_lse=attn_softmax_lse, ) output = output_tmp output_lse = output_lse_tmp return output, output_lse def _forward_prefill( self, q: torch.Tensor, kv_c_normed: torch.Tensor, k: torch.Tensor, kv_c_and_k_pe_cache: torch.Tensor, kv_c_and_k_pe_cache_scale: torch.Tensor, attn_metadata: MLACommonMetadata, ) -> torch.Tensor: assert attn_metadata.prefill is not None has_context = attn_metadata.prefill.chunked_context is not None kv_nope = self.kv_b_proj(kv_c_normed)[0].view(-1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim) k_nope, v_nope = kv_nope.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) v = v_nope k[...,:self.qk_nope_head_dim] = k_nope attn_output = torch.empty(q.shape[0], self.num_heads, self.v_head_dim, dtype=q.dtype, device=q.device) output = self.flash_attn_varlen_func( q=q, k=k, v=v, cu_seqlens_q=attn_metadata.prefill.query_start_loc, cu_seqlens_k=attn_metadata.prefill.query_start_loc, max_seqlen_q=attn_metadata.prefill.max_query_len, max_seqlen_k=attn_metadata.prefill.max_query_len, softmax_scale=self.scale, causal=True, return_softmax_lse=has_context, out=attn_output, ) if has_context: suffix_output, suffix_lse = output context_output, context_lse = self._compute_prefill_context( \ q, kv_c_and_k_pe_cache, kv_c_and_k_pe_cache_scale,attn_metadata) output = torch.empty_like(suffix_output) output = merge_attn_states( output=output, prefix_output=context_output, prefix_lse=context_lse, suffix_output=suffix_output, suffix_lse=suffix_lse, ) # slice by `:v.shape[-1]` in order to remove v headdim padding output = output.view(-1, self.num_heads * self.v_head_dim) return self.o_proj(output)[0] @abstractmethod def _forward_decode( self, ql_nope: torch.Tensor, q_pe: torch.Tensor, kv_c_and_k_pe_cache: torch.Tensor, attn_metadata: M, ) -> torch.Tensor: raise NotImplementedError def forward( self, layer: AttentionLayer, hidden_states_or_q_c: torch.Tensor, # query in unified attn k_c_normed: torch.Tensor, # key in unified attn k_pe: torch.Tensor, # value in unified attn kv_cache: torch.Tensor, kv_cache_scale:torch.Tensor, attn_metadata: M, output: Optional[torch.Tensor] = None, ) -> torch.Tensor: assert output is not None, "Output tensor must be provided." if attn_metadata is None: # Profiling run. output = torch.empty(output.shape[0], self.o_proj.output_size,device=hidden_states_or_q_c.device, dtype=hidden_states_or_q_c.dtype) return output # num_actual_toks = attn_metadata.num_actual_tokens # # Inputs and outputs may be padded for CUDA graphs # output_padded = output # output = output[:num_actual_toks, ...] # hidden_states_or_q_c = hidden_states_or_q_c[:num_actual_toks, ...] # k_c_normed = k_c_normed[:num_actual_toks, ...] # k_pe = k_pe[:num_actual_toks, ...] # Restore head dim (for rotary embedding) k_pe = k_pe.unsqueeze(1) assert attn_metadata.num_decodes is not None and \ attn_metadata.num_prefills is not None and \ attn_metadata.num_decode_tokens is not None has_decode = attn_metadata.num_decodes > 0 has_prefill = attn_metadata.num_prefills > 0 num_decode_tokens = attn_metadata.num_decode_tokens decode_hs_or_q_c = hidden_states_or_q_c[:num_decode_tokens] decode_k_pe = k_pe[:num_decode_tokens] prefill_hs_or_q_c = hidden_states_or_q_c[num_decode_tokens:] prefill_k_pe = k_pe[num_decode_tokens:] prefill_k_c_normed = k_c_normed[num_decode_tokens:] if has_decode: assert attn_metadata.decode is not None decode_ql_nope, decode_q_pe = self._q_proj_and_k_up_proj(decode_hs_or_q_c) decode_q_pe, decode_k_pe = self.rotary_emb(attn_metadata.decode.input_positions, decode_q_pe, decode_k_pe) if has_prefill: assert attn_metadata.prefill is not None prefill_q = self.q_proj(prefill_hs_or_q_c)[0].view(-1, self.num_heads, self.qk_head_dim) prefill_q_pe = prefill_q[..., self.qk_nope_head_dim:] prefill_k = torch.empty_like(prefill_q) ixf_ops.mla_rope(attn_metadata.prefill.input_positions, prefill_q_pe, prefill_k_pe.squeeze(1), prefill_k[...,self.qk_nope_head_dim:], self.rotary_emb.cos_sin_cache) # write the latent and rope to kv cache write_kv_cache = (None, None) if kv_cache.numel() > 0: # TODO.. if has_decode: if envs.VLLM_USE_INT8_MLA: write_kv_cache = (None, None) ops.concat_and_cache_mla_int8( k_c_normed[:num_decode_tokens], decode_k_pe, kv_cache, kv_cache_scale, attn_metadata.slot_mapping.flatten()[:num_decode_tokens], kv_cache_dtype=self.kv_cache_dtype, scale=layer._k_scale, ) else: write_kv_cache = (k_c_normed[:num_decode_tokens], decode_k_pe) # ops.concat_and_cache_mla( # k_c_normed[:num_decode_tokens], # decode_k_pe, # kv_cache, # attn_metadata.slot_mapping.flatten()[:num_decode_tokens], # kv_cache_dtype=self.kv_cache_dtype, # scale=layer._k_scale, # ) if has_prefill: if envs.VLLM_USE_INT8_MLA: ops.concat_and_cache_mla_int8( prefill_k_c_normed, prefill_k[...,self.qk_nope_head_dim:], kv_cache, kv_cache_scale, attn_metadata.slot_mapping.flatten()[num_decode_tokens:], kv_cache_dtype=self.kv_cache_dtype, scale=layer._k_scale, ) else: ops.concat_and_cache_mla( prefill_k_c_normed, prefill_k[...,self.qk_nope_head_dim:], kv_cache, attn_metadata.slot_mapping.flatten()[num_decode_tokens:], kv_cache_dtype=self.kv_cache_dtype, scale=layer._k_scale, ) output = torch.empty(output.shape[0], self.o_proj.output_size, device=hidden_states_or_q_c.device, dtype=hidden_states_or_q_c.dtype) if has_prefill: output[num_decode_tokens:] = self._forward_prefill( prefill_q, prefill_k_c_normed, prefill_k, kv_cache,kv_cache_scale, attn_metadata) if has_decode: output[:num_decode_tokens] = self._forward_decode( decode_ql_nope, decode_q_pe, kv_cache, kv_cache_scale,attn_metadata, *write_kv_cache) return output