Llama 3 8B · Grouped-query attention

32 independent transformer blocks

Grouped-query attention

learnedself.attention(self.attention_norm(x), start_pos, freqs_cis, mask)

Queries use position-aware keys to retrieve context through shared value heads.

H=AV_{\mathrm G}

\Delta_{\mathrm{attn}}=HW_O

All 13 steps

{\htmlClass{vl-stream-q}{Q}}=xW_Q

{\htmlClass{vl-stream-k}{K}}=xW_K

{\htmlClass{vl-stream-v}{V}}=xW_V

\begin{bmatrix}z_{2j}^{\prime}\\z_{2j+1}^{\prime}\end{bmatrix}=\begin{bmatrix}\cos(t\omega_j)&-\sin(t\omega_j)\\\sin(t\omega_j)&\cos(t\omega_j)\end{bmatrix}\begin{bmatrix}z_{2j}\\z_{2j+1}\end{bmatrix}

\begin{bmatrix}z_{2j}^{\prime}\\z_{2j+1}^{\prime}\end{bmatrix}=\begin{bmatrix}\cos(t\omega_j)&-\sin(t\omega_j)\\\sin(t\omega_j)&\cos(t\omega_j)\end{bmatrix}\begin{bmatrix}z_{2j}\\z_{2j+1}\end{bmatrix}

{\htmlClass{vl-stream-k}{K_{\mathrm G}}}[h]={\htmlClass{vl-stream-k}{K_{\mathrm r}}}\big[\lfloor h/4\rfloor\big]

{\htmlClass{vl-stream-v}{V_{\mathrm G}}}[h]={\htmlClass{vl-stream-v}{V}}\big[\lfloor h/4\rfloor\big]

S={\htmlClass{vl-stream-q}{Q_{\mathrm r}}}{\htmlClass{vl-stream-k}{K_{\mathrm G}^{\mathsf T}}}

S_{\mathrm s}=S/\sqrt{128}

S_{\mathrm m}=S_{\mathrm s}+{\htmlClass{vl-stream-context}{M}}

A[b,i,t,s]=\frac{\exp(S_{\mathrm m}[b,i,t,s])}{\sum_{u=0}^{T-1}\exp(S_{\mathrm m}[b,i,t,u])}

H=AV_{\mathrm G}

H=\operatorname{reshape}(H)

\Delta_{\mathrm{attn}}=HW_O

in: $x_{\mathrm{attn}}$ [B batch, T positions, 4096 d] $R$ [T positions, 64 pairs] $M$ [T queries, T keys]
out: $\Delta_{\mathrm{attn}}$ [B batch, T positions, 4096 d]
parameters: 41,943,040 · × 32 blocks = 1,342,177,280

Parameters

41,943,040 parameters per instance, across its learned components

Exact per-instance accounting
Component	Calculation	Parameters
Query projection	$4{,}096 \times 4{,}096$	16,777,216
Key projection	4,096 × (8 × 128)	4,194,304
Value projection	4,096 × (8 × 128)	4,194,304
Output projection	$4{,}096 \times 4{,}096$	16,777,216
Per-instance total		41,943,040

Across 32 independent blocks: $41{,}943{,}040 \times 32 = 1{,}342{,}177{,}280$ parameters.

Derived from the pinned catalogue dimensions and implementation. Checkpoint tensors have not been independently counted.

Codemodel.py 65–296 · sku_list.py 19, 188–207

model.py 93–120

        self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
        model_parallel_size = fs_init.get_model_parallel_world_size()
        self.n_local_heads = args.n_heads // model_parallel_size
        self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
        self.n_rep = self.n_local_heads // self.n_local_kv_heads
        self.head_dim = args.dim // args.n_heads

        self.wq = ColumnParallelLinear(
            args.dim,
            args.n_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        )
        self.wk = ColumnParallelLinear(
            args.dim,
            self.n_kv_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        )
        self.wv = ColumnParallelLinear(
            args.dim,
            self.n_kv_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        )

model.py 153–160

        bsz, seqlen, _ = x.shape
        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)

        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)

        xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)

model.py 65–75

def apply_rotary_emb(
    xq: torch.Tensor,
    xk: torch.Tensor,
    freqs_cis: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
    return xq_out.type_as(xq), xk_out.type_as(xk)

model.py 156–160

        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)

        xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)

model.py 78–87

def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
    """torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
    bs, slen, n_kv_heads, head_dim = x.shape
    if n_rep == 1:
        return x
    return (
        x[:, :, :, None, :]
        .expand(bs, slen, n_kv_heads, n_rep, head_dim)
        .reshape(bs, slen, n_kv_heads * n_rep, head_dim)
    )

model.py 93–98

        self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
        model_parallel_size = fs_init.get_model_parallel_world_size()
        self.n_local_heads = args.n_heads // model_parallel_size
        self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
        self.n_rep = self.n_local_heads // self.n_local_kv_heads
        self.head_dim = args.dim // args.n_heads

model.py 171–183

        # repeat k/v heads if n_kv_heads < n_heads
        keys = repeat_kv(
            keys, self.n_rep
        )  # (bs, cache_len + seqlen, n_local_heads, head_dim)
        values = repeat_kv(
            values, self.n_rep
        )  # (bs, cache_len + seqlen, n_local_heads, head_dim)

        xq = xq.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
        keys = keys.transpose(1, 2)  # (bs, n_local_heads, cache_len + seqlen, head_dim)
        values = values.transpose(
            1, 2
        )  # (bs, n_local_heads, cache_len + seqlen, head_dim)

model.py 179–187

        xq = xq.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
        keys = keys.transpose(1, 2)  # (bs, n_local_heads, cache_len + seqlen, head_dim)
        values = values.transpose(
            1, 2
        )  # (bs, n_local_heads, cache_len + seqlen, head_dim)
        scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
        if mask is not None:
            scores = scores + mask  # (bs, n_local_heads, seqlen, cache_len + seqlen)
        scores = F.softmax(scores.float(), dim=-1).type_as(xq)

model.py 284–296

        mask = None
        if seqlen > 1:
            mask = torch.full((seqlen, seqlen), float("-inf"), device=tokens.device)

            mask = torch.triu(mask, diagonal=1)

            # When performing key-value caching, we compute the attention scores
            # only for the new sequence. Thus, the matrix of scores is of size
            # (seqlen, cache_len + seqlen), and the only masked entries are (i, j) for
            # j > cache_len + i, since row i corresponds to token cache_len + i.
            mask = torch.hstack(
                [torch.zeros((seqlen, start_pos), device=tokens.device), mask]
            ).type_as(h)

model.py 121–127

        self.wo = RowParallelLinear(
            args.n_heads * self.head_dim,
            args.dim,
            bias=False,
            input_is_parallel=True,
            init_method=lambda x: x,
        )

model.py 184–190

        scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
        if mask is not None:
            scores = scores + mask  # (bs, n_local_heads, seqlen, cache_len + seqlen)
        scores = F.softmax(scores.float(), dim=-1).type_as(xq)
        output = torch.matmul(scores, values)  # (bs, n_local_heads, seqlen, head_dim)
        output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
        return self.wo(output)

model.py 236–248

        self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
        self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)

    def forward(
        self,
        x: torch.Tensor,
        start_pos: int,
        freqs_cis: torch.Tensor,
        mask: Optional[torch.Tensor],
    ):
        h = x + self.attention(self.attention_norm(x), start_pos, freqs_cis, mask)
        out = h + self.feed_forward(self.ffn_norm(h))
        return out

sku_list.py 19

LLAMA3_VOCAB_SIZE = 128256

sku_list.py 188–207

def llama3_base_models() -> list[Model]:
    return [
        Model(
            core_model_id=CoreModelId.llama3_8b,
            description="Llama 3 8b model",
            huggingface_repo="meta-llama/Llama-3-8B",
            arch_args={
                "dim": 4096,
                "n_layers": 32,
                "n_heads": 32,
                "n_kv_heads": 8,
                "vocab_size": LLAMA3_VOCAB_SIZE,
                "ffn_dim_multiplier": 1.3,
                "multiple_of": 1024,
                "norm_eps": 1e-05,
                "rope_theta": 500000.0,
                "use_scaled_rope": False,
            },
            pth_file_count=1,
        ),

Sources7

Direct evidence

Meta’s original Llama 3 8B base catalogue entry supplies dimensions, head counts, vocabulary, MLP rounding settings, epsilon and unscaled RoPE settings.

config · meta-llama/llama-models / models/sku_list.py
Pinned revision 0e0b8c519242 ↗ · lines 19–207

Read code excerpt

LLAMA3_VOCAB_SIZE = 128256

… lines 20–187 omitted …

def llama3_base_models() -> list[Model]:
    return [
        Model(
            core_model_id=CoreModelId.llama3_8b,
            description="Llama 3 8b model",
            huggingface_repo="meta-llama/Llama-3-8B",
            arch_args={
                "dim": 4096,
                "n_layers": 32,
                "n_heads": 32,
                "n_kv_heads": 8,
                "vocab_size": LLAMA3_VOCAB_SIZE,
                "ffn_dim_multiplier": 1.3,
                "multiple_of": 1024,
                "norm_eps": 1e-05,
                "rope_theta": 500000.0,
                "use_scaled_rope": False,
            },
            pth_file_count=1,
        ),

Direct evidence

Separate bias-free Q/K/V projections use configured query and KV head widths, then reshape into heads.

qkv · meta-llama/llama3 / llama/model.py
Pinned revision a0940f9cf706 ↗ · lines 93–160

Read code excerpt

        self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
        model_parallel_size = fs_init.get_model_parallel_world_size()
        self.n_local_heads = args.n_heads // model_parallel_size
        self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
        self.n_rep = self.n_local_heads // self.n_local_kv_heads
        self.head_dim = args.dim // args.n_heads

        self.wq = ColumnParallelLinear(
            args.dim,
            args.n_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        )
        self.wk = ColumnParallelLinear(
            args.dim,
            self.n_kv_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        )
        self.wv = ColumnParallelLinear(
            args.dim,
            self.n_kv_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        )

… lines 121–152 omitted …

        bsz, seqlen, _ = x.shape
        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)

        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)

        xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)

Direct evidence

Complex-pair rotations apply to queries and keys, with their original dtype restored.

rope · meta-llama/llama3 / llama/model.py
Pinned revision a0940f9cf706 ↗ · lines 65–160

Read code excerpt

def apply_rotary_emb(
    xq: torch.Tensor,
    xk: torch.Tensor,
    freqs_cis: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
    return xq_out.type_as(xq), xk_out.type_as(xk)

… lines 76–155 omitted …

        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)

        xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)

Direct evidence

Each KV head serves a group of query heads; the reference implementation expands KV heads before score/value multiplication.

gqa · meta-llama/llama3 / llama/model.py
Pinned revision a0940f9cf706 ↗ · lines 78–183

Read code excerpt

def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
    """torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
    bs, slen, n_kv_heads, head_dim = x.shape
    if n_rep == 1:
        return x
    return (
        x[:, :, :, None, :]
        .expand(bs, slen, n_kv_heads, n_rep, head_dim)
        .reshape(bs, slen, n_kv_heads * n_rep, head_dim)
    )

… lines 88–92 omitted …

        self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
        model_parallel_size = fs_init.get_model_parallel_world_size()
        self.n_local_heads = args.n_heads // model_parallel_size
        self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
        self.n_rep = self.n_local_heads // self.n_local_kv_heads
        self.head_dim = args.dim // args.n_heads

… lines 99–170 omitted …

        # repeat k/v heads if n_kv_heads < n_heads
        keys = repeat_kv(
            keys, self.n_rep
        )  # (bs, cache_len + seqlen, n_local_heads, head_dim)
        values = repeat_kv(
            values, self.n_rep
        )  # (bs, cache_len + seqlen, n_local_heads, head_dim)

        xq = xq.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
        keys = keys.transpose(1, 2)  # (bs, n_local_heads, cache_len + seqlen, head_dim)
        values = values.transpose(
            1, 2
        )  # (bs, n_local_heads, cache_len + seqlen, head_dim)

Direct evidence

Queries multiply transposed keys, scores are divided by the square root of head width, a causal mask is added, and softmax runs over key positions.

scores · meta-llama/llama3 / llama/model.py
Pinned revision a0940f9cf706 ↗ · lines 179–296

Read code excerpt

        xq = xq.transpose(1, 2)  # (bs, n_local_heads, seqlen, head_dim)
        keys = keys.transpose(1, 2)  # (bs, n_local_heads, cache_len + seqlen, head_dim)
        values = values.transpose(
            1, 2
        )  # (bs, n_local_heads, cache_len + seqlen, head_dim)
        scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
        if mask is not None:
            scores = scores + mask  # (bs, n_local_heads, seqlen, cache_len + seqlen)
        scores = F.softmax(scores.float(), dim=-1).type_as(xq)

… lines 188–283 omitted …

        mask = None
        if seqlen > 1:
            mask = torch.full((seqlen, seqlen), float("-inf"), device=tokens.device)

            mask = torch.triu(mask, diagonal=1)

            # When performing key-value caching, we compute the attention scores
            # only for the new sequence. Thus, the matrix of scores is of size
            # (seqlen, cache_len + seqlen), and the only masked entries are (i, j) for
            # j > cache_len + i, since row i corresponds to token cache_len + i.
            mask = torch.hstack(
                [torch.zeros((seqlen, start_pos), device=tokens.device), mask]
            ).type_as(h)

Direct evidence

Normalized attention weights multiply values; head outputs are merged and passed through a learned bias-free output projection.

retrieval · meta-llama/llama3 / llama/model.py
Pinned revision a0940f9cf706 ↗ · lines 121–190

Read code excerpt

        self.wo = RowParallelLinear(
            args.n_heads * self.head_dim,
            args.dim,
            bias=False,
            input_is_parallel=True,
            init_method=lambda x: x,
        )

… lines 128–183 omitted …

        scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
        if mask is not None:
            scores = scores + mask  # (bs, n_local_heads, seqlen, cache_len + seqlen)
        scores = F.softmax(scores.float(), dim=-1).type_as(xq)
        output = torch.matmul(scores, values)  # (bs, n_local_heads, seqlen, head_dim)
        output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
        return self.wo(output)

Direct evidence

Attention receives attention_norm(x) and its update is added to x; the MLP receives ffn_norm(h) and its update is added to h.

handoff · meta-llama/llama3 / llama/model.py
Pinned revision a0940f9cf706 ↗ · lines 236–248

Read code excerpt

        self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
        self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)

    def forward(
        self,
        x: torch.Tensor,
        start_pos: int,
        freqs_cis: torch.Tensor,
        mask: Optional[torch.Tensor],
    ):
        h = x + self.attention(self.attention_norm(x), start_pos, freqs_cis, mask)
        out = h + self.feed_forward(self.ffn_norm(h))
        return out