- in
- [B batch, 64 heads, 1 query, 1001 keys] [B batch, 64 heads, 1001 cached positions, 128 value]
- out
- [B batch, 1 position, 7168 d]
- parameters
- 58,720,256 · × 60 MoE blocks = 3,523,215,360
- mode
- decode · P = 1,000 cached positions · L = P + 1 = 1,001 keys read
Parameters
58,720,256 parameters per instance, across its learned components
| Component | Calculation | Parameters |
|---|---|---|
| Output projection | 58,720,256 | |
| Per-instance total | 58,720,256 | |
Across 60 independent MoE blocks: parameters.
Derived from the pinned configuration and implementation, and equal to the count of the checkpoint’s tensors read from its safetensors headers (FP8 block scales and stored rotary buffers excluded).
Explanation
The output projection is the only step that mixes heads.
Codemodeling_deepseek.py 799–840, 848–852
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
query_states[:, :, :, self.qk_nope_head_dim :] = q_pe
key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
key_states[:, :, :, self.qk_nope_head_dim :] = k_pe
if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)
attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)
assert attention_mask is not None
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(
attn_weights, dim=-1, dtype=torch.float32
).to(query_states.dtype)
attn_weights = nn.functional.dropout(
attn_weights, p=self.attention_dropout, training=self.training
)
attn_output = torch.matmul(attn_weights, value_states) attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
attn_output = self.o_proj(attn_output)Sources2
Direct evidence
The rotary parts are rotated; each head’s query and key join their non-rotated and rotated parts, the key’s rotated part being the shared rotary key; scores are scaled, masked by the causal mask, normalized and used to mix values.
mla-prefill · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 799–840
Read code excerpt
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
query_states[:, :, :, self.qk_nope_head_dim :] = q_pe
key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
key_states[:, :, :, self.qk_nope_head_dim :] = k_pe
if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)
attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)
assert attention_mask is not None
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(
attn_weights, dim=-1, dtype=torch.float32
).to(query_states.dtype)
attn_weights = nn.functional.dropout(
attn_weights, p=self.attention_dropout, training=self.training
)
attn_output = torch.matmul(attn_weights, value_states)Direct evidence
Head outputs are flattened and projected back to the model width.
mla-output · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 848–852
Read code excerpt
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
attn_output = self.o_proj(attn_output)