DeepSeek V4.1-Flash 登陆昇腾：框架支持与 Engram 主机卸载

DeepSeek V4.1-Flash 已合入 main PR #16925 · commit 0deca3181 · 2026-09-21 合入 · 73 文件 +10554/-665 · 代码基线 a71b766ce

一、背景：DeepSeek 发了个新模型，昇腾为什么跑不动

二、前史：这个需求为什么是第二次提交

三、诉求与方案：在 A3 上完整跑起来，五件套补齐

四、实现：先看执行链路，再读三段代码

    roles.append(
        DeepseekV41LayerRole(
            layer_idx=layer_idx,
            compress_ratio=ratio,
            kv_source=...,  # ↳ 该层全局 KV 来自哪个源层（CED：解码器 KV 由编码器源层投影）
            index_source=...,  # ↳ 该层稀疏索引来自哪个索引源层（跨层共享 TopK）
            is_candidate_source=layer_idx == candidate_source,  # ⚡分层索引器：候选池只由首个 Full 层构建
            # Consumer layers inherit the selection policy of their index
            # source.  For example, layer 26 reuses layer 24 TopK, and that
            # TopK was computed inside layer 20's candidate blocks.  ← 原注释
            uses_candidate_filter=index_source is not None and index_source > candidate_source,
            engram_slot=engram_slots.get(layer_idx),  # ↳ 该层是否挂 Engram 条件记忆注入
        )
    )

def get_layer_tuples(specs):
    """Return DSV4-style ordered layer tuples and their physical page sizes."""
    mla = {name for name, spec in specs.items() if isinstance(spec, AscendMLAAttentionSpec)}
    state = sorted((name for name, spec in specs.items() if isinstance(spec, CircularBufferSpec)), key=_layer_number)
    swa = {name for name, spec in specs.items() if isinstance(spec, AscendSlidingWindowMLASpec)}
    ...
    for slot_idx, kv_name in enumerate(full):
        prefix = kv_name.rsplit(".", 1)[0]
        index_name = prefix + ".indexer.k_cache"  # ⚡索引器 K 缓存与全局 KV 同槽配对
        index_spec = specs[index_name]
        kv_spec = specs[kv_name]
        aliases = ([state[slot_idx]] if slot_idx < len(state) else []) + target_swa[slot_idx :: len(full)]
        kv_bytes = kv_spec.unpadded_page_size_bytes
        index_bytes = index_spec.unpadded_page_size_bytes
        ...
        capacity = max(  # ⚡页容量=槽内最大成员：索引 K 与 FP16 scale 排在 KV 之后同页（压缩比2：64 行/131072B；压缩比1：128 行/147712B）
            kv_bytes + index_bytes,
            *(specs[name].unpadded_page_size_bytes for name in aliases),
        )
        layer_tuples.append((kv_name, index_name, *aliases))  # ⚡一个块 ID 同时定位四类缓存（层外置）
        page_sizes.append(capacity)
    return page_sizes, layer_tuples

def forward(self, attn, positions, hidden_states, output: torch.Tensor | None = None):
    # ↳ 自定义算子调用方传入图稳定输出缓冲：O 投影结果直写，免去图边界的整张拷贝（⚡昇腾图优化）
    if output is None:
        output = torch.empty_like(hidden_states)
    forward_context = get_forward_context()
    if forward_context.attn_metadata is None:
        output.zero_()  # ↳ dummy 前向（剖析/捕获期）直接清零返回
        return output
    metadata = self._get_layer_metadata(forward_context.attn_metadata)
    self._prepare_inputs_and_caches(attn, hidden_states, metadata, forward_context.attn_metadata)
    num_tokens = metadata.swa.num_actual_tokens
    if num_tokens:
        positions = metadata.positions[:num_tokens]
        cos, sin = metadata.rope(attn.rotary_emb.layername, num_tokens)
        q, qr = self._prepare_queries(attn, hidden_states, positions, cos, sin, metadata)  # ⚡阶段①：双流 Q 投影∥KV Cube
        compressed_indices = self._select_sparse_indices(attn, hidden_states, qr, positions, cos, sin, metadata)  # ⚡阶段②：QLI V3 候选三模式
        attention_output = self._forward_attention(attn, q, metadata, compressed_indices)  # →NPU 阶段③：SparseFlashMla 双阶段
        torch.ops._C_ascend.inplace_partial_rotary_mul(  # →NPU：输出侧逆 RoPE 原地完成，不落临时张量
            attention_output.unsqueeze(1),
            cos,
            -sin,
            rotary_mode="interleave",
            partial_slice=[attn.nope_head_dim, attn.head_dim],
        )
    else:
        heads = attn.n_heads if getattr(attn, "enable_dsa_cp", False) else attn.n_local_heads
        attention_output = hidden_states.new_empty((0, heads, attn.head_dim))
    self._project_output(attn, attention_output, hidden_states, metadata, projected=output)
    return output

五、后续：三个跟进 PR 与部署形态

六、相关链接