Skip to content

LLM inference in C/C++

C++ ggml

model : Granite-Switch Architecture (#25107)

Bar Haim157b81fe6dbf
diff --git a/conversion/__init__.py b/conversion/__init__.py
index 1f781a7903..3b8bebdaea 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -103,6 +103,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
"GraniteMoeForCausalLM": "granite",
"GraniteMoeHybridForCausalLM": "granite",
"GraniteMoeSharedForCausalLM": "granite",
+ "GraniteSwitchForCausalLM": "granite",
"GraniteSpeechForConditionalGeneration": "granite",
"GraniteSpeechPlusForConditionalGeneration": "granite",
"Grok1ForCausalLM": "grok",
diff --git a/conversion/granite.py b/conversion/granite.py
index 8367ed225d..956342e6d6 100644
--- a/conversion/granite.py
+++ b/conversion/granite.py
@@ -123,6 +123,166 @@ class GraniteMoeModel(GraniteModel):
yield from super().modify_tensors(data_torch, name, bid)
[email protected]("GraniteSwitchForCausalLM")
+class GraniteSwitchModel(GraniteMoeModel):
+ """Dense, all-attention Granite with N per-token embedded LoRA adapters, stacked
+ over the adapter dim with a zero adapter at slot 0 (N = num_adapters + 1)."""
+ model_arch = gguf.MODEL_ARCH.GRANITE_SWITCH
+
+ # permute q/k per-slice below (NORM-rope layout), not via the parent's auto-permute
+ undo_permute = False
+
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+ # the weightless switch reserves one cache slot: one fewer block than num_hidden_layers
+ self.block_count = self.block_count - 1
+ self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+ self._n_adapters = int(self.hparams["num_adapters"])
+ self._max_lora_rank = int(self.hparams["max_lora_rank"])
+ self._n_slots = self._n_adapters + 1 # +1 for the zero slot at index 0
+
+ n_head = int(self.hparams["num_attention_heads"])
+ n_kv_head = int(self.hparams["num_key_value_heads"])
+ head_dim = (
+ self.hparams.get("projection_head_dim")
+ or self.hparams.get("head_dim")
+ or (self.hparams["hidden_size"] // n_head)
+ )
+ self._n_head = n_head
+ self._n_kv_head = n_kv_head
+ self._head_dim = int(head_dim)
+ self._q_size = n_head * self._head_dim
+ self._kv_size = n_kv_head * self._head_dim
+
+ def set_gguf_parameters(self):
+ super().set_gguf_parameters()
+
+ # dense: pin expert_used_count to 0 (config carries a leftover num_experts_per_tok)
+ if not self.hparams.get("num_local_experts"):
+ self.gguf_writer.add_expert_used_count(0)
+
+ self.gguf_writer.add_adapter_count(self._n_adapters)
+ self.gguf_writer.add_adapter_lora_rank(self._max_lora_rank)
+ self.gguf_writer.add_adapter_token_ids_activate(self.hparams["adapter_token_ids"])
+ self.gguf_writer.add_adapter_token_ids_substitute(self.hparams["adapter_substitute_token_ids"])
+ router_gain = float(self.hparams.get("control_token_gain", 15.0))
+ self.gguf_writer.add_adapter_router_gain(router_gain)
+ logger.info("gguf: (graniteswitch) num_adapters=%s max_lora_rank=%s n_slots=%s router_gain=%s", self._n_adapters, self._max_lora_rank, self._n_slots, router_gain)
+
+ def _lora_a(self, data: Tensor) -> Tensor:
+ # on-disk A: [n_adapters, 1, max_rank, in] -> [n_adapters+1, max_rank, in]
+ a = data.squeeze(1)
+ zero = torch.zeros_like(a[:1])
+ return torch.cat([zero, a], dim=0).contiguous()
+
+ def _lora_b(self, data: Tensor, permute_n_head: int | None = None) -> Tensor:
+ # on-disk B: [n_adapters, 1, out, max_rank] -> [n_adapters+1, out, max_rank]
+ b = data.squeeze(1)
+ if permute_n_head is not None:
+ # permute each adapter's B output rows to match the permuted q/k base
+ b = torch.stack([self.permute(b[i], permute_n_head, permute_n_head) for i in range(b.shape[0])], dim=0)
+ zero = torch.zeros_like(b[:1])
+ return torch.cat([zero, b], dim=0).contiguous()
+
+ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+ T = gguf.MODEL_TENSOR
+
+ # skip the weightless switch + control-token buffers (rebuilt at load time)
+ bare = name.split(".")[-1]
+ if (
+ name.startswith("model.switch.") or name.startswith("switch.")
+ or bare in ("adapter_token_ids", "control_to_substitute_lut")
+ ):
+ return
+
+ if "self_attn.qkv_proj" in name:
+ if name.endswith("base_layer.weight"):
+ # fused [q|k|v] rows: permute q/k row-blocks for ggml's NORM-rope layout
+ q, k, v = data_torch.split([self._q_size, self._kv_size, self._kv_size], dim=0)
+ q = self.permute(q, self._n_head, self._n_head)
+ k = self.permute(k, self._n_kv_head, self._n_kv_head)
+ fused = torch.cat([q, k, v], dim=0)
+ yield (self.format_tensor_name(T.ATTN_QKV, bid), fused)
+ return
+ if "lora_A_slices." in name:
+ slot = int(name.rsplit(".", 1)[1])
+ key = {0: T.ATTN_Q, 1: T.ATTN_K, 2: T.ATTN_V}[slot]
+ yield (self.format_tensor_name(key, bid, suffix=".lora_a"), self._lora_a(data_torch))
+ return
+ if "lora_B_slices." in name:
+ slot = int(name.rsplit(".", 1)[1])
+ key, ph = {
+ 0: (T.ATTN_Q, self._n_head),
+ 1: (T.ATTN_K, self._n_kv_head),
+ 2: (T.ATTN_V, None),
+ }[slot]
+ yield (self.format_tensor_name(key, bid, suffix=".lora_b"), self._lora_b(data_torch, ph))
+ return
+ raise ValueError(f"Unexpected qkv_proj tensor: {name}")
+
+ if "self_attn.o_proj" in name:
+ if name.endswith("base_layer.weight"):
+ yield (self.format_tensor_name(T.ATTN_OUT, bid), data_torch)
+ return
+ if name.endswith("lora_A"):
+ yield (self.format_tensor_name(T.ATTN_OUT, bid, suffix=".lora_a"), self._lora_a(data_torch))
+ return
+ if name.endswith("lora_B"):
+ yield (self.format_tensor_name(T.ATTN_OUT, bid, suffix=".lora_b"), self._lora_b(data_torch))
+ return
+ raise ValueError(f"Unexpected o_proj tensor: {name}")
+
+ if "shared_mlp.input_linear" in name:
+ ffn = self.hparams["shared_intermediate_size"]
+ if name.endswith("base_layer.weight"):
+ gate, up = data_torch.split([ffn, ffn], dim=0)
+ yield (self.format_tensor_name(T.FFN_GATE, bid), gate)
+ yield (self.format_tensor_name(T.FFN_UP, bid), up)
+ return
+ if "lora_A_slices." in name:
+ slot = int(name.rsplit(".", 1)[1])
+ key = {0: T.FFN_GATE, 1: T.FFN_UP}[slot]
+ yield (self.format_tensor_name(key, bid, suffix=".lora_a"), self._lora_a(data_torch))
+ return
+ if "lora_B_slices." in name:
+ slot = int(name.rsplit(".", 1)[1])
+ key = {0: T.FFN_GATE, 1: T.FFN_UP}[slot]
+ yield (self.format_tensor_name(key, bid, suffix=".lora_b"), self._lora_b(data_torch))
+ return
+ raise ValueError(f"Unexpected shared_mlp.input_linear tensor: {name}")
+
+ if "shared_mlp.output_linear" in name:
+ if name.endswith("base_layer.weight"):
+ yield (self.format_tensor_name(T.FFN_DOWN, bid), data_torch)
+ return
+ if name.endswith("lora_A"):
+ yield (self.format_tensor_name(T.FFN_DOWN, bid, suffix=".lora_a"), self._lora_a(data_torch))
+ return
+ if name.endswith("lora_B"):
+ yield (self.format_tensor_name(T.FFN_DOWN, bid, suffix=".lora_b"), self._lora_b(data_torch))
+ return
+ raise ValueError(f"Unexpected shared_mlp.output_linear tensor: {name}")
+
+ if bid is not None and ".layers." in name and (
+ "input_layernorm" in name or "post_attention_layernorm" in name
+ ):
+ key = T.ATTN_NORM if "input_layernorm" in name else T.FFN_NORM
+ yield (self.format_tensor_name(key, bid), data_torch)
+ return
+
+ if name in ("model.embed_tokens.weight", "embed_tokens.weight"):
+ yield (self.format_tensor_name(T.TOKEN_EMBD), data_torch)
+ return
+ if name in ("model.norm.weight", "norm.weight"):
+ yield (self.format_tensor_name(T.OUTPUT_NORM), data_torch)
+ return
+ if name == "lm_head.weight":
+ return # tied to token_embd
+
+ raise ValueError(f"graniteswitch: unhandled tensor {name!r} (bid={bid})")
+
+
@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
"""GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 8516222ccc..304100f7f9 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -164,6 +164,13 @@ class Keys:
NORM_BEFORE_RESIDUAL = "{arch}.norm_before_residual"
NORM_BEFORE_FC = "{arch}.norm_before_fc"
+ class Adapters:
+ COUNT = "{arch}.adapters.count"
+ TOKEN_IDS_ACTIVATE = "{arch}.adapters.token_ids_activate"
+ TOKEN_IDS_SUBSTITUTE = "{arch}.adapters.token_ids_substitute"
+ LORA_RANK = "{arch}.adapters.lora_rank"
+ ROUTER_GAIN = "{arch}.adapters.router_gain"
+
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
HEAD_COUNT_KV = "{arch}.attention.head_count_kv"
@@ -527,6 +534,7 @@ class MODEL_ARCH(IntEnum):
GRANITE = auto()
GRANITE_MOE = auto()
GRANITE_HYBRID = auto()
+ GRANITE_SWITCH = auto()
CHAMELEON = auto()
WAVTOKENIZER_DEC = auto()
PLM = auto()
@@ -1198,6 +1206,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.GRANITE: "granite",
MODEL_ARCH.GRANITE_MOE: "granitemoe",
MODEL_ARCH.GRANITE_HYBRID: "granitehybrid",
+ MODEL_ARCH.GRANITE_SWITCH: "graniteswitch",
MODEL_ARCH.CHAMELEON: "chameleon",
MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
MODEL_ARCH.PLM: "plm",
@@ -3972,6 +3981,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
+ MODEL_ARCH.GRANITE_SWITCH: [
+ MODEL_TENSOR.TOKEN_EMBD,
+ MODEL_TENSOR.OUTPUT_NORM,
+ MODEL_TENSOR.OUTPUT,
+ MODEL_TENSOR.ATTN_NORM,
+ MODEL_TENSOR.ATTN_QKV,
+ MODEL_TENSOR.ATTN_Q,
+ MODEL_TENSOR.ATTN_K,
+ MODEL_TENSOR.ATTN_V,
+ MODEL_TENSOR.ATTN_OUT,
+ MODEL_TENSOR.FFN_NORM,
+ MODEL_TENSOR.FFN_GATE,
+ MODEL_TENSOR.FFN_DOWN,
+ MODEL_TENSOR.FFN_UP,
+ ],
MODEL_ARCH.CHAMELEON: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index 39da9f2c05..81ae07c11a 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -906,6 +906,21 @@ class GGUFWriter:
def add_embedding_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
+ def add_adapter_count(self, count: int) -> None:
+ self.add_uint32(Keys.Adapters.COUNT.format(arch=self.arch), count)
+
+ def add_adapter_token_ids_activate(self, ids: Sequence[int]) -> None:
+ self.add_array(Keys.Adapters.TOKEN_IDS_ACTIVATE.format(arch=self.arch), ids)
+
+ def add_adapter_token_ids_substitute(self, ids: Sequence[int]) -> None:
+ self.add_array(Keys.Adapters.TOKEN_IDS_SUBSTITUTE.format(arch=self.arch), ids)
+
+ def add_adapter_lora_rank(self, rank: int) -> None:
+ self.add_uint32(Keys.Adapters.LORA_RANK.format(arch=self.arch), rank)
+
+ def add_adapter_router_gain(self, gain: float) -> None:
+ self.add_float32(Keys.Adapters.ROUTER_GAIN.format(arch=self.arch), gain)
+
def add_wkv_head_size(self, size: int) -> None:
self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 836cfade22..b8231ba6f8 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -100,6 +100,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GRANITE, "granite" },
{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
{ LLM_ARCH_GRANITE_HYBRID, "granitehybrid" },
+ { LLM_ARCH_GRANITE_SWITCH, "graniteswitch" },
{ LLM_ARCH_CHAMELEON, "chameleon" },
{ LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
{ LLM_ARCH_PLM, "plm" },
@@ -220,6 +221,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_TIME_DECAY_EXTRA_DIM, "%s.time_decay_extra_dim" },
{ LLM_KV_RESIDUAL_SCALE, "%s.residual_scale" },
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
+ { LLM_KV_ADAPTER_COUNT, "%s.adapters.count" },
+ { LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE, "%s.adapters.token_ids_activate" },
+ { LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE, "%s.adapters.token_ids_substitute" },
+ { LLM_KV_ADAPTER_LORA_RANK, "%s.adapters.lora_rank" },
+ { LLM_KV_ADAPTER_ROUTER_GAIN, "%s.adapters.router_gain" },
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
{ LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
{ LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" },
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 49c2a6ac39..47adc3d684 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -105,6 +105,7 @@ enum llm_arch {
LLM_ARCH_GRANITE,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_GRANITE_HYBRID,
+ LLM_ARCH_GRANITE_SWITCH,
LLM_ARCH_CHAMELEON,
LLM_ARCH_WAVTOKENIZER_DEC,
LLM_ARCH_PLM,
@@ -225,6 +226,11 @@ enum llm_kv {
LLM_KV_TIME_DECAY_EXTRA_DIM,
LLM_KV_RESIDUAL_SCALE,
LLM_KV_EMBEDDING_SCALE,
+ LLM_KV_ADAPTER_COUNT,
+ LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE,
+ LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE,
+ LLM_KV_ADAPTER_LORA_RANK,
+ LLM_KV_ADAPTER_ROUTER_GAIN,
LLM_KV_TOKEN_SHIFT_COUNT,
LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
LLM_KV_FULL_ATTENTION_INTERVAL,
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 19cca7df1e..6f2bf1362c 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -3602,8 +3602,9 @@ llama_context * llama_init_from_model(
model->hparams.pooling_type, params.pooling_type);
}
+ // router_layer >= 0 means n_layer_nextn is repurposed for a router layer, not real MTP
if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP &&
- model->hparams.n_layer_nextn == 0) {
+ (model->hparams.n_layer_nextn == 0 || model->hparams.router_layer >= 0)) {
LLAMA_LOG_WARN("%s: context type MTP requested but model doesn't contain MTP layers\n", __func__);
return nullptr;
}
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index 846d4c69a6..781277f3ff 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -277,6 +277,16 @@ bool llama_hparams::has_kv(uint32_t il) const {
return true;
}
+bool llama_hparams::has_rope(uint32_t il) const {
+ // the router layer stores adapter routing signal, not positional info,
+ // so it must not be RoPE-shifted
+ if (router_layer >= 0 && (int32_t) il == router_layer) {
+ return false;
+ }
+
+ return true;
+}
+
uint32_t llama_hparams::n_layer() const {
return n_layer_all - n_layer_nextn;
}
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 6e8336c987..57de808242 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -53,6 +53,10 @@ struct llama_hparams {
uint32_t n_embd;
uint32_t n_layer_all;
uint32_t n_layer_nextn = 0;
+
+ // granite-switch: index of the single-head "router" KV layer that encodes
+ // per-token adapter selection. -1 when the model has no such layer.
+ int32_t router_layer = -1;
uint32_t n_expert = 0;
uint32_t n_expert_used = 0;
uint32_t n_rel_attn_bkts = 0;
@@ -371,6 +375,8 @@ struct llama_hparams {
bool has_kv(uint32_t il) const;
+ bool has_rope(uint32_t il) const;
+
// number of effective layers (excludes nextn layers)
uint32_t n_layer() const;
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 8678a326d9..5382cd7266 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -1931,6 +1931,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
for (const auto & layer : layers) {
const uint32_t il = layer.il;
+ if (!hparams.has_rope(il)) {
+ continue;
+ }
+
const int64_t n_head_kv = hparams.n_head_kv(il);
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 71bc9f7ef0..3d50f8a1cb 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -937,10 +937,11 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
} break;
case GGML_OP_MUL_MAT_ID:
{
- const int n_expert_used = hparams.n_expert_used;
- GGML_ASSERT(n_expert_used > 0);
- ggml_tensor * b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_expert_used, 512);
- ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_expert_used, 512);
+ // Used for either MoE expert routing or embedded adapter routing
+ const int n_ids_used = hparams.router_layer >= 0 ? 1 : hparams.n_expert_used;
+ GGML_ASSERT(n_ids_used > 0);
+ ggml_tensor * b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_ids_used, 512);
+ ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_ids_used, 512);
op_tensor = ggml_mul_mat_id(ctx, w, b, ids);
} break;
case GGML_OP_ADD:
@@ -1123,15 +1124,14 @@ struct ggml_tensor * llama_model_loader::create_tensor(
return nullptr;
}
- // tensors with "bias" suffix are always used with GGML_OP_ADD or GGML_OP_ADD_ID
+ // tensors with "bias" suffix are always used with GGML_OP_ADD or GGML_OP_ADD_ID;
+ // embedded-adapter ".lora_a"/".lora_b" tensors are always used with GGML_OP_MUL_MAT_ID
ggml_op op;
- bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0;
- if (bias) {
- if (info.op == GGML_OP_MUL_MAT_ID) {
- op = GGML_OP_ADD_ID;
- } else {
- op = GGML_OP_ADD;
- }
+ if (tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0) {
+ op = info.op == GGML_OP_MUL_MAT_ID ? GGML_OP_ADD_ID : GGML_OP_ADD;
+ } else if (hparams.router_layer >= 0 && tn.suffix != nullptr &&
+ (strcmp(tn.suffix, "lora_a") == 0 || strcmp(tn.suffix, "lora_b") == 0)) {
+ op = GGML_OP_MUL_MAT_ID;
} else {
op = info.op;
}
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 4cc1c0a1c2..b4575b82a1 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -234,6 +234,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_granite(params);
case LLM_ARCH_GRANITE_MOE:
return new llama_model_granite_moe(params);
+ case LLM_ARCH_GRANITE_SWITCH:
+ return new llama_model_granite_switch(params);
case LLM_ARCH_MINICPM:
return new llama_model_minicpm(params);
case LLM_ARCH_GRANITE_HYBRID:
@@ -1912,6 +1914,7 @@ void llama_model::print_info() const {
arch == LLM_ARCH_GRANITE ||
arch == LLM_ARCH_GRANITE_MOE ||
arch == LLM_ARCH_GRANITE_HYBRID ||
+ arch == LLM_ARCH_GRANITE_SWITCH ||
arch == LLM_ARCH_NEMOTRON_H_MOE) {
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
@@ -2596,6 +2599,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
case LLM_ARCH_GRANITE_HYBRID:
+ case LLM_ARCH_GRANITE_SWITCH:
case LLM_ARCH_CHAMELEON:
case LLM_ARCH_BAILINGMOE:
case LLM_ARCH_NEO_BERT:
diff --git a/src/llama-model.h b/src/llama-model.h
index 6b9e94a0a6..1dd0904387 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -223,6 +223,24 @@ struct llama_layer_nextn {
struct ggml_tensor * shared_head_norm = nullptr;
};
+struct llama_layer_switch_lora {
+ struct ggml_tensor * a_q = nullptr;
+ struct ggml_tensor * b_q = nullptr;
+ struct ggml_tensor * a_k = nullptr;
+ struct ggml_tensor * b_k = nullptr;
+ struct ggml_tensor * a_v = nullptr;
+ struct ggml_tensor * b_v = nullptr;
+ struct ggml_tensor * a_o = nullptr;
+ struct ggml_tensor * b_o = nullptr;
+
+ struct ggml_tensor * a_gate = nullptr;
+ struct ggml_tensor * b_gate = nullptr;
+ struct ggml_tensor * a_up = nullptr;
+ struct ggml_tensor * b_up = nullptr;
+ struct ggml_tensor * a_down = nullptr;
+ struct ggml_tensor * b_down = nullptr;
+};
+
struct llama_layer {
// normalization
struct ggml_tensor * attn_norm = nullptr;
@@ -533,6 +551,8 @@ struct llama_layer {
struct llama_layer_shortconv shortconv;
struct llama_layer_nextn nextn;
+
+ struct llama_layer_switch_lora switch_lora;
};
struct llama_device {
diff --git a/src/models/granite-switch.cpp b/src/models/granite-switch.cpp
new file mode 100644
index 0000000000..80f6b86edc
--- /dev/null
+++ b/src/models/granite-switch.cpp
@@ -0,0 +1,426 @@
+#include "models.h"
+
+#include <cmath>
+
+void llama_model_granite_switch::load_arch_hparams(llama_model_loader & ml) {
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+ ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
+ ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale, false);
+ ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, false);
+ ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale, false);
+
+ bool rope_finetuned = true;
+ ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+ hparams.rope_finetuned = rope_finetuned;
+
+ switch (hparams.n_layer()) {
+ case 40: type = hparams.n_embd == 4096 ? LLM_TYPE_8B : LLM_TYPE_3B; break;
+ case 64: type = LLM_TYPE_30B; break;
+ default: type = LLM_TYPE_UNKNOWN;
+ }
+
+ ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+
+ ml.get_key(LLM_KV_ADAPTER_COUNT, n_adapters);
+ ml.get_key(LLM_KV_ADAPTER_LORA_RANK, max_lora_rank);
+ ml.get_key(LLM_KV_ADAPTER_ROUTER_GAIN, router_gain, /* required */ false);
+
+ // bound counts that size tensors
+ if (n_adapters > 4096) {
+ throw std::runtime_error(format("graniteswitch: invalid adapter count %u", n_adapters));
+ }
+ if (max_lora_rank > 4096) {
+ throw std::runtime_error(format("graniteswitch: invalid lora rank %u", max_lora_rank));
+ }
+
+ std::vector<llama_token> token_ids;
+ std::vector<llama_token> substitute_ids;
+ ml.get_arr(LLM_KV_ADAPTER_TOKEN_IDS_ACTIVATE, token_ids);
+ ml.get_arr(LLM_KV_ADAPTER_TOKEN_IDS_SUBSTITUTE, substitute_ids);
+
+ if (token_ids.size() != n_adapters || substitute_ids.size() != n_adapters) {
+ throw std::runtime_error(format(
+ "graniteswitch: adapter token id arrays (%zu activate, %zu substitute) do not match adapter count %u",
+ token_ids.size(), substitute_ids.size(), n_adapters));
+ }
+
+ adapter_token_to_slot.clear();
+ adapter_token_to_substitute.clear();
+ for (uint32_t i = 0; i < n_adapters; ++i) {
+ // adapter i -> stacked slot i+1 (slot 0 is the base/zero delta)
+ adapter_token_to_slot[token_ids[i]] = (int32_t) (i + 1);
+ adapter_token_to_substitute[token_ids[i]] = substitute_ids[i];
+ }
+
+ // extra single-head attention layer at the END (index n_real) holds the router
+ // K/V. reusing n_layer_nextn keeps n_layer() == n_real, so the regular layers
+ // keep their indices and the KV cache shift/defrag skips the router layer.
+ // n_layer_nextn is repurposed here (no MTP): it leaks as 1 into the
+ // llama_model_n_layer_nextn() getter and a re-saved nextn_predict_layers
+ const uint32_t n_real = hparams.n_layer();
+ if (n_real >= LLAMA_MAX_LAYERS) {
+ throw std::runtime_error(format("graniteswitch: block count %u exceeds LLAMA_MAX_LAYERS", n_real));
+ }
+ hparams.router_layer = (int32_t) n_real;
+ hparams.n_layer_all = n_real + 1;
+ hparams.n_layer_nextn = 1;
+
+ hparams.n_head_arr[n_real] = 1;
+ hparams.n_head_kv_arr[n_real] = 1;
+ hparams.n_ff_arr[n_real] = 0;
+}
+
+void llama_model_granite_switch::load_arch_tensors(llama_model_loader &) {
+ LLAMA_LOAD_LOCALS;
+
+ const int64_t n_slots = (int64_t) n_adapters + 1; // slot 0 = base/zero delta
+ const int64_t n_rank = (int64_t) max_lora_rank;
+ const int64_t n_embd_q = n_embd_head_k * n_head;
+ const int64_t n_embd_kv = n_embd_k_gqa;
+
+ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+ // substitute ids index tok_embd rows directly; range-check against n_vocab
+ for (const auto & kv : adapter_token_to_substitute) {
+ const llama_token sub = kv.second;
+ if (sub < 0 || (int64_t) sub >= n_vocab) {
+ throw std::runtime_error(format(
+ "graniteswitch: substitute token id %d out of range [0, %d)", sub, (int) n_vocab));
+ }
+ }
+
+ output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+ output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+ if (output == NULL) {
+ output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+ }
+
+ for (int i = 0; i < n_layer; ++i) {
+ auto & layer = layers[i];
+
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+ layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd_q + 2*n_embd_kv}, 0);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q, n_embd}, 0);
+
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+ layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+ layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
+ layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
+
+ auto & sl = layer.switch_lora;
+
+ sl.a_q = create_tensor(tn(LLM_TENSOR_ATTN_Q, "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+ sl.b_q = create_tensor(tn(LLM_TENSOR_ATTN_Q, "lora_b", i), {n_rank, n_embd_q, n_slots}, 0);
+ sl.a_k = create_tensor(tn(LLM_TENSOR_ATTN_K, "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+ sl.b_k = create_tensor(tn(LLM_TENSOR_ATTN_K, "lora_b", i), {n_rank, n_embd_kv, n_slots}, 0);
+ sl.a_v = create_tensor(tn(LLM_TENSOR_ATTN_V, "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+ sl.b_v = create_tensor(tn(LLM_TENSOR_ATTN_V, "lora_b", i), {n_rank, n_embd_kv, n_slots}, 0);
+
+ sl.a_o = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "lora_a", i), {n_embd_q, n_rank, n_slots}, 0);
+ sl.b_o = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "lora_b", i), {n_rank, n_embd, n_slots}, 0);
+
+ sl.a_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+ sl.b_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "lora_b", i), {n_rank, n_ff, n_slots}, 0);
+ sl.a_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "lora_a", i), {n_embd, n_rank, n_slots}, 0);
+ sl.b_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "lora_b", i), {n_rank, n_ff, n_slots}, 0);
+ sl.a_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "lora_a", i), { n_ff, n_rank, n_slots}, 0);
+ sl.b_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "lora_b", i), {n_rank, n_embd, n_slots}, 0);
+ }
+}
+
+class llm_graph_input_switch : public llm_graph_input_i {
+public:
+ llm_graph_input_switch(const llama_model_granite_switch & smodel) : smodel(smodel) {}
+ virtual ~llm_graph_input_switch() = default;
+
+ void set_input(const llama_ubatch * ubatch) override;
+
+ ggml_tensor * sub_tokens = nullptr; // I32 [n_tokens] adapter-substituted token ids
+ ggml_tensor * router_ksig = nullptr; // F32 [n_tokens] router K signal (+/-gain)
+ ggml_tensor * router_vval = nullptr; // F32 [n_tokens] router V value (adapter slot / 0)
+ ggml_tensor * router_q = nullptr; // F32 [n_tokens] router Q value (constant 1.0)
+
+ const llama_model_granite_switch & smodel;
+};
+
+// K dim-0 is +gain for an adapter token, -gain otherwise; the causal softmax then
+// lets a single visible adapter token dominate so the readback recovers its slot.
+void llm_graph_input_switch::set_input(const llama_ubatch * ubatch) {
+ if (!ubatch->token) {
+ return;
+ }
+
+ const int64_t n_tokens = ubatch->n_tokens;
+
+ std::vector<int32_t> sub (n_tokens);
+ std::vector<float> ksig(n_tokens);
+ std::vector<float> vval(n_tokens);
+ std::vector<float> q (n_tokens, 1.0f);
+
+ for (int64_t i = 0; i < n_tokens; ++i) {
+ const llama_token tok = ubatch->token[i];
+
+ const auto it = smodel.adapter_token_to_slot.find(tok);
+ if (it != smodel.adapter_token_to_slot.end()) {
+ ksig[i] = +smodel.router_gain;
+ vval[i] = (float) it->second;
+ } else {
+ ksig[i] = -smodel.router_gain;
+ vval[i] = 0.0f;
+ }
+
+ const auto sit = smodel.adapter_token_to_substitute.find(tok);
+ sub[i] = (sit != smodel.adapter_token_to_substitute.end())
+ ? (int32_t) sit->second
+ : (int32_t) tok;
+ }
+
+ ggml_backend_tensor_set(sub_tokens, sub.data(), 0, n_tokens*ggml_element_size(sub_tokens));
+ ggml_backend_tensor_set(router_ksig, ksig.data(), 0, n_tokens*ggml_element_size(router_ksig));
+ ggml_backend_tensor_set(router_vval, vval.data(), 0, n_tokens*ggml_element_size(router_vval));
+ ggml_backend_tensor_set(router_q, q.data(), 0, n_tokens*ggml_element_size(router_q));
+}
+
+std::unique_ptr<llm_graph_context> llama_model_granite_switch::build_arch_graph(const llm_graph_params & params) const {
+ return std::make_unique<graph>(*this, params);
+}
+
+// per-token switched LoRA delta: B_a*(A_a*x), adapter selected per token via ids.
+// cur: {n_in, n_tokens}, ids: {n_tokens} -> {n_out, n_tokens}
+ggml_tensor * llama_model_granite_switch::graph::build_switched_lora_delta(
+ ggml_tensor * lora_a,
+ ggml_tensor * lora_b,
+ ggml_tensor * cur,
+ ggml_tensor * ids) {
+ const int64_t n_in = cur->ne[0];
+ const int64_t n_tokens = cur->ne[1];
+
+ ggml_tensor * x = ggml_reshape_3d(ctx0, cur, n_in, 1, n_tokens);
+ ggml_tensor * ids2 = ggml_reshape_2d(ctx0, ids, 1, n_tokens);
+
+ ggml_tensor * a = ggml_mul_mat_id(ctx0, lora_a, x, ids2); // {max_rank, 1, n_tokens}
+ ggml_tensor * d = ggml_mul_mat_id(ctx0, lora_b, a, ids2); // {n_out, 1, n_tokens}
+
+ return ggml_reshape_2d(ctx0, d, d->ne[0], n_tokens);
+}
+
+ggml_tensor * llama_model_granite_switch::graph::build_switched_lora_mm(
+ ggml_tensor * w,
+ ggml_tensor * lora_a,
+ ggml_tensor * lora_b,
+ ggml_tensor * cur,
+ ggml_tensor * ids) {
+ ggml_tensor * base = ggml_mul_mat(ctx0, w, cur);
+ ggml_tensor * delta = build_switched_lora_delta(lora_a, lora_b, cur, ids);
+ return ggml_add(ctx0, base, delta);
+}
+
+llama_model_granite_switch::graph::graph(
+ const llama_model & model,
+ const llm_graph_params & params)
+ : llm_graph_context(params) {
+
+ const auto & smodel = static_cast<const llama_model_granite_switch &>(model);
+
+ // TODO: support raw embedding input (multimodal / pre-embedded tokens) when needed
+ GGML_ASSERT(ubatch.token && "granite-switch requires token input");
+
+ const int64_t n_embd_head = hparams.n_embd_head_v();
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+ GGML_ASSERT(n_embd_head == n_rot);
+
+ auto inp_switch = std::make_unique<llm_graph_input_switch>(smodel);
+ inp_switch->sub_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+ inp_switch->router_ksig = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
+ inp_switch->router_vval = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
+ inp_switch->router_q = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
+ ggml_set_input(inp_switch->sub_tokens);
+ ggml_set_input(inp_switch->router_ksig);
+ ggml_set_input(inp_switch->router_vval);
+ ggml_set_input(inp_switch->router_q);
+ ggml_tensor * sub_tokens = inp_switch->sub_tokens;
+ ggml_tensor * router_ksig = inp_switch->router_ksig;
+ ggml_tensor * router_vval = inp_switch->router_vval;
+ ggml_tensor * router_q = inp_switch->router_q;
+ res->add_input(std::move(inp_switch));
+
+ // embed the substituted ids directly; build_inp_embd would embed the raw tokens
+ ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embd, sub_tokens);
+ if (hparams.f_embedding_scale != 0.0f) {
+ inpL = ggml_scale(ctx0, inpL, hparams.f_embedding_scale);
+ }
+ cb(inpL, "inp_embd", -1);
+
+ ggml_tensor * inp_pos = nullptr;
+ if (hparams.rope_finetuned) {
+ inp_pos = build_inp_pos();
+ }
+ auto * inp_attn = build_attn_inp_kv();
+
+ // single causal head at layer R recovers the adapter index in-graph: only dim 0
+ // carries signal (Q[0]=1, K[0]=+/-gain, V[0]=slot/0), the rest is zero-padded.
+ const int R = hparams.router_layer;
+ GGML_ASSERT(R >= 0);
+ auto router_lane = [&](ggml_tensor * sig1d) {
+ ggml_tensor * t = ggml_reshape_3d(ctx0, sig1d, 1, 1, n_tokens);
+ return ggml_pad(ctx0, t, (int) n_embd_head - 1, 0, 0, 0);
+ };
+ ggml_tensor * Qr = router_lane(router_q);
+ ggml_tensor * Kr = router_lane(router_ksig);
+ ggml_tensor * Vr = router_lane(router_vval);
+
+ ggml_tensor * router_out = build_attn(inp_attn,
+ nullptr, nullptr, nullptr,
+ Qr, Kr, Vr, nullptr, nullptr, nullptr, /*kq_scale=*/1.0f, /*il=*/R);
+ cb(router_out, "router_out", R);
+
+ // row 0 of router_out is the attended slot; clamp+round to an I32 index
+ ggml_tensor * slot_f = ggml_cont(ctx0,
+ ggml_view_2d(ctx0, router_out, 1, n_tokens, router_out->nb[1], 0));
+ slot_f = ggml_reshape_1d(ctx0, slot_f, n_tokens);
+ slot_f = ggml_clamp(ctx0, slot_f, 0.0f, (float) smodel.n_adapters);
+ slot_f = ggml_round(ctx0, slot_f);
+ ggml_tensor * adapter_ids = ggml_cast(ctx0, slot_f, GGML_TYPE_I32);
+ cb(adapter_ids, "adapter_ids", -1);
+
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ ggml_tensor * cur;
+
+ for (int il = 0; il < n_layer; ++il) {
+ ggml_tensor * inpSA = inpL;
+
+ cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+ cb(cur, "attn_norm", il);
+
+ cur = build_attention_layer(cur, inp_pos, adapter_ids, inp_attn, model, n_embd_head, il);
+
+ if (il == n_layer - 1 && inp_out_ids) {
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+ inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+ // keep adapter_ids aligned to the kept rows (2D round-trip for get_rows)
+ const int64_t n_out = inp_out_ids->ne[0];
+ adapter_ids = ggml_get_rows(ctx0,
+ ggml_reshape_2d(ctx0, adapter_ids, 1, adapter_ids->ne[0]), inp_out_ids);
+ adapter_ids = ggml_reshape_1d(ctx0, adapter_ids, n_out);
+ }
+
+ cur = build_layer_ffn(cur, inpSA, adapter_ids, model, il);
+
+ inpL = cur;
+ }
+
+ cur = inpL;
+
+ cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+ cb(cur, "result_norm", -1);
+ res->t_embd = cur;
+
+ cur = build_lora_mm(model.output, cur, model.output_s);
+
+ cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+ cb(cur, "result_output", -1);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llama_model_granite_switch::graph::build_attention_layer(
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ ggml_tensor * adapter_ids,
+ llm_graph_input_attn_kv * inp_attn,
+ const llama_model & model,
+ const int64_t n_embd_head,
+ const int il) {
+
+ const auto & layer = model.layers[il];
+ const auto & sl = layer.switch_lora;
+
+ const int64_t n_head = hparams.n_head(il);
+ const int64_t n_head_kv = hparams.n_head_kv(il);
+
+ ggml_tensor * qkv = ggml_mul_mat(ctx0, layer.wqkv, cur);
+ cb(qkv, "wqkv", il);
+
+ const int64_t n_embd_q = n_embd_head * n_head;
+ const int64_t n_embd_kv = n_embd_head * n_head_kv;
+
+ // slice fused qkv into Q/K/V, made contiguous so LoRA deltas can be added
+ ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_q, qkv->ne[1], qkv->nb[1], 0));
+ ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_kv, qkv->ne[1], qkv->nb[1], n_embd_q*ggml_element_size(qkv)));
+ ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_kv, qkv->ne[1], qkv->nb[1], (n_embd_q + n_embd_kv)*ggml_element_size(qkv)));
+
+ Qcur = ggml_add(ctx0, Qcur, build_switched_lora_delta(sl.a_q, sl.b_q, cur, adapter_ids));
+ Kcur = ggml_add(ctx0, Kcur, build_switched_lora_delta(sl.a_k, sl.b_k, cur, adapter_ids));
+ Vcur = ggml_add(ctx0, Vcur, build_switched_lora_delta(sl.a_v, sl.b_v, cur, adapter_ids));
+
+ Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+ Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+ Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+ if (hparams.rope_finetuned) {
+ ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+ Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow);
+ }
+ cb(Qcur, "Qcur", il);
+ cb(Kcur, "Kcur", il);
+ cb(Vcur, "Vcur", il);
+
+ const float kq_scale = hparams.f_attention_scale == 0.0f
+ ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+ // wo = nullptr so build_attn returns concatenated heads; o-proj is switched below
+ ggml_tensor * attn = build_attn(inp_attn,
+ nullptr, nullptr, nullptr,
+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+ cb(attn, "attn_pre_o", il);
+
+ cur = build_switched_lora_mm(layer.wo, sl.a_o, sl.b_o, attn, adapter_ids);
+ cb(cur, "attn_out", il);
+ return cur;
+}
+
+ggml_tensor * llama_model_granite_switch::graph::build_layer_ffn(
+ ggml_tensor * cur,
+ ggml_tensor * inpSA,
+ ggml_tensor * adapter_ids,
+ const llama_model & model,
+ const int il) {
+
+ const auto & layer = model.layers[il];
+ const auto & sl = layer.switch_lora;
+
+ if (hparams.f_residual_scale) {
+ cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+ }
+ ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+ cb(ffn_inp, "ffn_inp", il);
+
+ cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
+
+ ggml_tensor * g = build_switched_lora_mm(layer.ffn_gate, sl.a_gate, sl.b_gate, cur, adapter_ids);
+ ggml_tensor * u = build_switched_lora_mm(layer.ffn_up, sl.a_up, sl.b_up, cur, adapter_ids);
+ g = ggml_silu(ctx0, g);
+ ggml_tensor * gu = ggml_mul(ctx0, g, u);
+ cur = build_switched_lora_mm(layer.ffn_down, sl.a_down, sl.b_down, gu, adapter_ids);
+ cb(cur, "ffn_out", il);
+
+ if (hparams.f_residual_scale) {
+ cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+ }
+ cur = ggml_add(ctx0, cur, ffn_inp);
+
+ cur = build_cvec(cur, il);
+ cb(cur, "l_out", il);
+
+ return cur;
+}
diff --git a/src/models/models.h b/src/models/models.h
index ad3dadaf39..a8908da429 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -1596,6 +1596,56 @@ struct llama_model_granite_moe : public llama_model_base {
};
+struct llama_model_granite_switch : public llama_model_base {
+ llama_model_granite_switch(const struct llama_model_params & params) : llama_model_base(params) {}
+ void load_arch_hparams(llama_model_loader & ml) override;
+ void load_arch_tensors(llama_model_loader & ml) override;
+
+ uint32_t n_adapters = 0;
+ uint32_t max_lora_rank = 0;
+ float router_gain = 15.0f;
+
+ std::unordered_map<llama_token, int32_t> adapter_token_to_slot;
+ std::unordered_map<llama_token, llama_token> adapter_token_to_substitute;
+
+ struct graph : public llm_graph_context {
+ graph(const llama_model & model, const llm_graph_params & params);
+
+ private:
+ ggml_tensor * build_switched_lora_delta(
+ ggml_tensor * lora_a,
+ ggml_tensor * lora_b,
+ ggml_tensor * cur,
+ ggml_tensor * ids);
+
+ ggml_tensor * build_switched_lora_mm(
+ ggml_tensor * w,
+ ggml_tensor * lora_a,
+ ggml_tensor * lora_b,
+ ggml_tensor * cur,
+ ggml_tensor * ids);
+
+ ggml_tensor * build_attention_layer(
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ ggml_tensor * adapter_ids,
+ llm_graph_input_attn_kv * inp_attn,
+ const llama_model & model,
+ const int64_t n_embd_head,
+ const int il);
+
+ ggml_tensor * build_layer_ffn(
+ ggml_tensor * cur,
+ ggml_tensor * inpSA,
+ ggml_tensor * adapter_ids,
+ const llama_model & model,
+ const int il);
+ };
+
+ std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
struct llama_model_minicpm : public llama_model_base {
llama_model_minicpm(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp
index 6ff0d0ac1d..d1a648c87e 100644
--- a/tests/test-llama-archs.cpp
+++ b/tests/test-llama-archs.cpp
@@ -411,6 +411,9 @@ static bool arch_supported(const llm_arch arch) {
if (arch == LLM_ARCH_GEMMA4 || arch == LLM_ARCH_GEMMA4_ASSISTANT) {
return false; // FIXME @ngxson
}
+ if (arch == LLM_ARCH_GRANITE_SWITCH) {
+ return false; // FIXME adapter fixture
+ }
if (arch == LLM_ARCH_LLAMA_EMBED || arch == LLM_ARCH_GEMMA_EMBEDDING || arch == LLM_ARCH_T5ENCODER) {
return false; // FIXME Embedding (?) models produce inconsistent results.
}