GLM-5 · Sparse attention · OpenModelAtlas

78 transformer blocks › 75 MoE blocks › Sparse latent attention

Sparse attention

learnedGlmMoeDsaAttention.forward · 445–461

Attend over each query’s selected positions with 64 heads and project the result to the model width.

O=AV

\Delta_{\mathrm{attn}}=oW_{O}

All 9 steps

\text{Attention weights:}

S=QK^{\mathsf T}

S'=\frac{S}{\sqrt{256}}

\tilde S=S'+{\htmlClass{vl-stream-context}{M^{\mathrm{sel}}}}

A[b,i,t,s]=\frac{\exp(\tilde S[b,i,t,s])}{\sum_{u=0}^{1000}\exp(\tilde S[b,i,t,u])}

\text{Mix and project:}

O=AV

o=\operatorname{reshape}(O)

\Delta_{\mathrm{attn}}=oW_{O}

in: $Q$ [B batch, 64 heads, 1 positions, 256 content + rotary] $K$ [B batch, 64 heads, 1001 cached positions, 256 content + rotary] $V$ [B batch, 64 heads, 1001 cached positions, 256 value] $M^{\mathrm{sel}}$ [B batch, 1 queries, 1001 keys]
out: $\Delta_{\mathrm{attn}}$ [B batch, 1 positions, 6144 d]
parameters: 100,663,296 · × 75 MoE blocks = 7,549,747,200

Parameters

100,663,296 parameters per instance, across its learned components

Exact per-instance accounting
Component	Calculation	Parameters
Mix and project	100,663,296 across its components	100,663,296
Per-instance total		100,663,296

Across 75 independent MoE blocks: $100{,}663{,}296 \times 75 = 7{,}549{,}747{,}200$ parameters.

Derived from the pinned configuration and transformers implementation, and equal to the count of the checkpoint’s tensors read from its safetensors headers, less the multi-token-prediction layer the implementation does not load. The total includes the routers’ 19,200 correction-bias values, which the implementation holds as buffers.

Explanation

These cards give the eager backend’s code, eager_attention_forward (modeling_glm_moe_dsa.py 268–290). Under SDPA (638) the selection is a boolean mask (438–439); the flash-mla kernel reads the indices directly (443, 456).

Codemodeling_glm_moe_dsa.py 268–462

modeling_glm_moe_dsa.py 268–290

def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights

modeling_glm_moe_dsa.py 404–458

        cos, sin = position_embeddings
        q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)

        # Cache read / write is performed while latent KV is still compressed
        if past_key_values is not None:
            k_pass, k_rot = past_key_values.update(k_pass, k_rot, self.layer_idx)

        query_states = torch.cat((q_pass, q_rot), dim=-1)

        key_states, value_states = self.expand_kv(k_pass, k_rot)

        # DSA: select this layer's top-k tokens, or reuse the previous full layer's on `"shared"` layers.
        if self.indexer is not None:
            topk_indices = self.indexer(
                hidden_states,
                q_resid,
                position_embeddings,
                attention_mask[:, 0, :, :],
                position_ids,  # Kept for BC
                past_key_values=past_key_values,
            )  # [B, S, topk]
        else:
            if prev_topk_indices is None:
                raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.")
            topk_indices = prev_topk_indices

        sparse_indices = None
        if self.config._attn_implementation in ("eager", "sdpa"):
            index_mask = (
                topk_indices.new_ones((batch_size, seq_length, key_states.shape[2]), dtype=torch.bool)
                .scatter(-1, topk_indices.long(), False)
                .unsqueeze(1)
            )

            if attention_mask.dtype == torch.bool:
                attention_mask = attention_mask & ~index_mask
            else:
                attention_mask = attention_mask.masked_fill(index_mask, torch.finfo(hidden_states.dtype).min)
        else:
            sparse_indices = topk_indices

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            indices=sparse_indices,  # consumed by flash_mla_with_kvcache; ignored by eager / SDPA
            **kwargs,
        )

modeling_glm_moe_dsa.py 460–462

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights, topk_indices

Sources2

Direct evidence

The rotary parts are rotated; the normalized latent and the rotated key are what the cache stores; each head joins its content and rotary parts; the indexer’s selection becomes a mask that sets unselected positions to the dtype’s most negative value on top of the causal mask; attention is scaled dot-product attention with a float32 softmax.

mla-prefill · huggingface/transformers / modeling_glm_moe_dsa.py
Pinned revision 5e4d6304de55 ↗ · lines 268–458

Read code excerpt

def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights

… lines 291–403 omitted …

        cos, sin = position_embeddings
        q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)

        # Cache read / write is performed while latent KV is still compressed
        if past_key_values is not None:
            k_pass, k_rot = past_key_values.update(k_pass, k_rot, self.layer_idx)

        query_states = torch.cat((q_pass, q_rot), dim=-1)

        key_states, value_states = self.expand_kv(k_pass, k_rot)

        # DSA: select this layer's top-k tokens, or reuse the previous full layer's on `"shared"` layers.
        if self.indexer is not None:
            topk_indices = self.indexer(
                hidden_states,
                q_resid,
                position_embeddings,
                attention_mask[:, 0, :, :],
                position_ids,  # Kept for BC
                past_key_values=past_key_values,
            )  # [B, S, topk]
        else:
            if prev_topk_indices is None:
                raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.")
            topk_indices = prev_topk_indices

        sparse_indices = None
        if self.config._attn_implementation in ("eager", "sdpa"):
            index_mask = (
                topk_indices.new_ones((batch_size, seq_length, key_states.shape[2]), dtype=torch.bool)
                .scatter(-1, topk_indices.long(), False)
                .unsqueeze(1)
            )

            if attention_mask.dtype == torch.bool:
                attention_mask = attention_mask & ~index_mask
            else:
                attention_mask = attention_mask.masked_fill(index_mask, torch.finfo(hidden_states.dtype).min)
        else:
            sparse_indices = topk_indices

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            indices=sparse_indices,  # consumed by flash_mla_with_kvcache; ignored by eager / SDPA
            **kwargs,
        )

Direct evidence

Head outputs are flattened and projected back to the model width.

mla-output · huggingface/transformers / modeling_glm_moe_dsa.py
Pinned revision 5e4d6304de55 ↗ · lines 460–462

Read code excerpt

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights, topk_indices