with open(dir_model / "config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+ self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+ self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
+ self.n_group = self.find_hparam(["n_groups"], optional=True) or 1
def set_vocab(self):
vocab_size = self.hparams["vocab_size"]
self._set_vocab_builtin("gpt-neox", vocab_size)
def set_gguf_parameters(self):
- d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
- d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
- d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model
- d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 128
- head_dim = self.find_hparam(["mamba_d_head", "head_dim"], optional=True) or 64
- n_group = self.find_hparam(["n_groups"], optional=True) or 1
+ d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
+ d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 128
+ head_dim = self.find_hparam(["mamba_d_head", "head_dim"], optional=True) or 64
rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
# TODO: does this really matter?
# skip the assertion for FalconH1 Model
if self.model_arch != gguf.MODEL_ARCH.FALCON_H1:
- assert d_inner == 2 * d_model
- assert d_inner % head_dim == 0
+ assert self.d_inner == 2 * self.d_model
+ assert self.d_inner % head_dim == 0
self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
- self.gguf_writer.add_embedding_length(d_model)
+ self.gguf_writer.add_embedding_length(self.d_model)
self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
self.gguf_writer.add_block_count(self.block_count)
self.gguf_writer.add_ssm_conv_kernel(d_conv)
- self.gguf_writer.add_ssm_inner_size(d_inner)
+ self.gguf_writer.add_ssm_inner_size(self.d_inner)
self.gguf_writer.add_ssm_state_size(d_state)
- self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
- self.gguf_writer.add_ssm_group_count(n_group)
+ self.gguf_writer.add_ssm_time_step_rank(self.d_inner // head_dim)
+ self.gguf_writer.add_ssm_group_count(self.n_group)
self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
self.gguf_writer.add_file_type(self.ftype)
# (D is also unsqueezed, but for more straightforward broadcast internally)
data_torch = data_torch.reshape((*data_torch.shape, 1))
elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
- d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
- d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model
- n_group = self.hparams.get("n_groups", 1)
- data_torch = data_torch.reshape((n_group, d_inner // n_group))
+ data_torch = data_torch.reshape((self.n_group, self.d_inner // self.n_group))
if name.endswith(".A_log"):
logger.debug("A_log --> A ==> " + new_name)
(self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
]
+ has_experts = bool(self.hparams.get('num_local_experts'))
+
if name.endswith("shared_mlp.input_linear.weight"):
ffn_dim = self.hparams["shared_intermediate_size"]
assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
gate, up = data_torch.split(ffn_dim, dim=-2)
+ if has_experts:
+ return [
+ (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+ (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+ ]
return [
- (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
- (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+ (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), gate),
+ (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), up),
+ ]
+
+ if not has_experts and name.endswith("shared_mlp.output_linear.weight"):
+ return [
+ (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), data_torch)
]
return super().modify_tensors(data_torch, name, bid)
+@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
+class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
+ """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
+ layers and optionally uses MoE w/ a shared expert"""
+ model_arch = gguf.MODEL_ARCH.GRANITE_HYBRID
+ undo_permute = True
+
+ def __init__(self, *args, **kwargs):
+
+ # Hybrid mamba models use a prefix for the mamba-specific params.
+ # TODO: Extend this if the prefix(es) need to be configurable
+ self.hparam_prefixes = ["mamba"]
+
+ super().__init__(*args, **kwargs)
+
+ # Lists of which layers use ssm vs attention
+ self._attn_layers = self.get_attn_layers()
+ self._ssm_layers = [
+ i for i in range(self.block_count)
+ if i not in self._attn_layers
+ ]
+
+ # n_group and d_inner are used during reshape_tensors for mamba2
+ self.d_model = self.find_hparam(["hidden_size", "d_model"])
+ self.n_group = self.find_hparam(["n_groups"])
+ self.d_inner = self.find_hparam(["expand"]) * self.d_model
+
+ def get_attn_layers(self):
+ # Explicit list of layer type names
+ if layer_types := self.hparams.get("layer_types"):
+ return [
+ i for i, typ in enumerate(layer_types)
+ if typ == "attention"
+ ]
+
+ # Layer types indicated by index or period
+ attn_layers = self.hparams.get("attn_layer_indices", [])
+ if not attn_layers:
+ attn_period = self.hparams.get("attn_layer_period")
+ assert attn_period, "Didn't find attn_layer_indices or attn_layer_period"
+ attn_offset = self.hparams.get("attn_layer_offset")
+ assert attn_offset is not None, "No attention layer offset set with attn_layer_period"
+ attn_layers = [
+ i for i in range(self.block_count)
+ if i % attn_period == attn_offset
+ ]
+ return attn_layers
+
+ def find_hparam(self, keys: Iterable[str], *args, **kwargs) -> Any:
+ prefixed = []
+ for pfx in self.hparam_prefixes:
+ prefixed.extend(
+ "_".join([pfx, k])
+ for k in keys
+ )
+ keys = list(keys) + prefixed
+ return Mamba2Model.find_hparam(self, keys, *args, **kwargs)
+
+ def modify_tensors(
+ self, data_torch: Tensor, name: str, bid: int | None
+ ) -> Iterable[tuple[str, Tensor]]:
+ if (
+ name.endswith("block_sparse_moe.input_linear.weight")
+ or "shared_mlp" in name
+ ):
+ return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+
+ # Determine whether this is a mamba layer or an attention layer
+ if bid in self._ssm_layers:
+ return Mamba2Model.modify_tensors(self, data_torch, name, bid)
+ elif bid in self._attn_layers:
+ return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+ return [(self.map_tensor_name(name), data_torch)]
+
+ def set_gguf_parameters(self):
+ """This method merges params from both parents and some that are
+ specific to this model. The result is some duplication of how the params
+ get set. The following warnings are expected during conversion:
+
+ WARNING:Duplicated key name 'granitehybrid.attention.head_count_kv'
+ WARNING:Duplicated key name 'granitehybrid.context_length'
+ """
+ GraniteMoeModel.set_gguf_parameters(self)
+
+ ## Mamba mixer params ##
+ self.gguf_writer.add_ssm_conv_kernel(self.find_hparam(["conv_kernel", "d_conv"]))
+ self.gguf_writer.add_ssm_state_size(self.find_hparam(["state_size", "d_state"]))
+ self.gguf_writer.add_ssm_group_count(self.n_group)
+ self.gguf_writer.add_ssm_inner_size(self.d_inner)
+ # NOTE: The mamba_dt_rank is _not_ the right field for how this is used
+ # in llama.cpp
+ self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads"]))
+
+ ## Attention params ##
+ head_count_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
+ head_count_kv_vec = [
+ head_count_kv if i in self._attn_layers else 0 for i in range(self.block_count)
+ ]
+ if rope_dim := self.hparams.get("attn_rotary_emb"):
+ self.gguf_writer.add_rope_dimension_count(rope_dim)
+ self.gguf_writer.add_head_count_kv(head_count_kv_vec)
+
+ ## If Bamba, use rope, otherwise don't
+ use_rope = "BambaForCausalLM" in self.hparams["architectures"]
+ self.gguf_writer.add_rope_scaling_finetuned(use_rope)
+ if not use_rope:
+ self.gguf_writer.add_context_length(2**20)
+
+ ## Validation ##
+ d_head = self.find_hparam(["d_head"], optional=True) or 64
+ assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported"
+ assert self.d_inner % d_head == 0, f"SSM inner size {self.d_inner} not a multiple of head dim {d_head}"
+
+ def set_vocab(self):
+ self.hparams["pad_vocab_size_multiple"] = 8
+ Mamba2Model.set_vocab(self)
+
+
@ModelBase.register("BailingMoeForCausalLM")
class BailingMoeModel(TextModel):
model_arch = gguf.MODEL_ARCH.BAILINGMOE
# Use Llama conversion for attention
self._transformer_model_class = LlamaModel
- # n_group and d_inner are used during reshape_tensors for mamaba2
+ # n_group and d_inner are used during reshape_tensors for mamba2
self.n_group = self.find_hparam(["n_groups"])
self.d_inner = self.find_hparam(["mamba_d_ssm"])
self.d_head = self.find_hparam(["d_head"])
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale);
+ // Granite uses rope_finetuned as a switch for rope, so default to true
+ 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 32: type = LLM_TYPE_3B; break;
case 40: type = LLM_TYPE_3B; break;
default: type = LLM_TYPE_UNKNOWN;
}
+ // For Granite MoE Shared
+ ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+ } break;
+ case LLM_ARCH_GRANITE_HYBRID:
+ {
+ 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, /* required */ false);
+ ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale, /* required */ false);
+ ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, /* required */ false);
+ ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale, /* required */ false);
+
+ ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv);
+ ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner);
+ ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state);
+ ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+ ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
+
+ // Granite uses rope_finetuned as a switch for rope, so default to true
+ bool rope_finetuned = true;
+ ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+ hparams.rope_finetuned = rope_finetuned;
+
+ // A layer is recurrent IFF the n_head_kv value is set to 0
+ for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+ hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
+ }
+
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+ switch (hparams.n_layer) {
+ // TODO: Add llm type label (not sure this is useful)
+ default: type = LLM_TYPE_UNKNOWN;
+ }
+
// For Granite MoE Shared
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
} break;
}
}
} break;
+ case LLM_ARCH_GRANITE_HYBRID:
+ {
+ // mamba2 Mixer SSM params
+ // NOTE: int64_t for tensor dimensions
+ const int64_t d_conv = hparams.ssm_d_conv;
+ const int64_t d_inner = hparams.ssm_d_inner;
+ const int64_t d_state = hparams.ssm_d_state;
+ const int64_t n_ssm_head = hparams.ssm_dt_rank;
+ const int64_t n_group = hparams.ssm_n_group;
+ const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+
+ // only an expansion factor of 2 is supported for now
+ GGML_ASSERT(2 * n_embd == d_inner);
+
+ // embeddings
+ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+ // output
+ {
+ 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 is NULL, init from the input tok embed, duplicated to allow offloading
+ 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];
+
+ // norm
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+ if (hparams.is_recurrent(i)) {
+ // ssm layers
+ layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+ layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+ layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);
+
+ layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
+
+ // no "weight" suffix for these
+ layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
+ layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
+
+ layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
+ // out_proj
+ layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+ } else {
+ // attention layers (with optional bias)
+ const int64_t n_head_i = hparams.n_head(i);
+ const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
+ const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
+ layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head_i}, 0);
+ layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa_i}, 0);
+ layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa_i}, 0);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
+ layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+ layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
+ layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
+ layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+ }
+
+ // feed forward (w/ optional biases)
+ if (n_expert > 0) {
+ // MoE FFN
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+ layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+ layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+ layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+ layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0);
+ layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
+
+ // For Granite MoE Shared
+ if (hparams.n_ff_shexp > 0) {
+ layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+ layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+ layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+ }
+ } else {
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+ layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 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);
+ layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+ layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+ layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+ }
+ }
+ } break;
case LLM_ARCH_XVERSE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
if (arch == LLM_ARCH_MAMBA ||
arch == LLM_ARCH_MAMBA2 ||
arch == LLM_ARCH_JAMBA ||
- arch == LLM_ARCH_FALCON_H1) {
+ arch == LLM_ARCH_FALCON_H1 ||
+ arch == LLM_ARCH_GRANITE_HYBRID) {
LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv);
LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner);
LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state);
if (arch == LLM_ARCH_MINICPM ||
arch == LLM_ARCH_GRANITE ||
- arch == LLM_ARCH_GRANITE_MOE) {
+ arch == LLM_ARCH_GRANITE_MOE ||
+ arch == LLM_ARCH_GRANITE_HYBRID) {
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);
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
}
};
-
struct llm_build_granite : public llm_graph_context {
llm_build_granite(
const llama_model & model,
const llm_graph_params & params,
- ggml_cgraph * gf,
- const bool use_rope = true)
+ ggml_cgraph * gf)
: llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
// inp_pos - built only if rope enabled
ggml_tensor * inp_pos = nullptr;
- if (use_rope) {
+ if (hparams.rope_finetuned) {
inp_pos = build_inp_pos();
}
auto * inp_attn = build_attn_inp_kv_unified();
- const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
cb(cur, "attn_norm", il);
// self-attention
- {
- // compute Q and K and (optionally) RoPE them
- ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
- cb(Qcur, "Qcur", il);
- if (model.layers[il].bq) {
- Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
- cb(Qcur, "Qcur", il);
- }
+ cur = build_attention_layer(
+ gf, cur, inp_pos, inp_attn,
+ model, n_embd_head, il);
- ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
- cb(Kcur, "Kcur", il);
- if (model.layers[il].bk) {
- Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
- cb(Kcur, "Kcur", 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);
+ }
- ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
- cb(Vcur, "Vcur", il);
- if (model.layers[il].bv) {
- Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
- cb(Vcur, "Vcur", il);
- }
+ // ffn
+ cur = build_layer_ffn(cur, inpSA, model, il);
- 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);
+ // input for next layer
+ inpL = cur;
+ }
- if (use_rope) {
- 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
- );
+ cur = inpL;
- 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
- );
- }
+ cur = build_norm(cur,
+ model.output_norm, NULL,
+ LLM_NORM_RMS, -1);
- cb(Qcur, "Qcur", il);
- cb(Kcur, "Kcur", il);
- cb(Vcur, "Vcur", il);
+ cb(cur, "result_norm", -1);
+ res->t_embd = cur;
- cur = build_attn(inp_attn, gf,
- model.layers[il].wo, model.layers[il].bo,
- Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+ // lm_head
+ cur = build_lora_mm(model.output, cur);
+
+ // For Granite architectures - scale logits
+ 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 * build_attention_layer(
+ ggml_cgraph * gf,
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ llm_graph_input_attn_kv_unified * inp_attn,
+ const llama_model & model,
+ const int64_t n_embd_head,
+ const int il) {
+
+ // compute Q and K and (optionally) RoPE them
+ ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+ cb(Qcur, "Qcur", il);
+ if (model.layers[il].bq) {
+ Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+ cb(Qcur, "Qcur", il);
+ }
+
+ ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+ cb(Kcur, "Kcur", il);
+ if (model.layers[il].bk) {
+ Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+ cb(Kcur, "Kcur", il);
+ }
+
+ ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+ cb(Vcur, "Vcur", il);
+ if (model.layers[il].bv) {
+ Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+ cb(Vcur, "Vcur", il);
+ }
+
+ Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
+ Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+ Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+ const bool use_rope = hparams.rope_finetuned;
+ if (use_rope) {
+ 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;
+ cur = build_attn(inp_attn, gf,
+ model.layers[il].wo, model.layers[il].bo,
+ Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
- }
+ return cur;
+ }
- 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);
- }
+ ggml_tensor * build_layer_ffn(
+ ggml_tensor * cur,
+ ggml_tensor * inpSA,
+ const llama_model & model,
+ const int il) {
- // For Granite architectures - scale residual
+ // For Granite architectures - scale residual
+ 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);
+ }
+ ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+ cb(ffn_inp, "ffn_inp", il);
- // feed-forward network (non-MoE)
- if (model.layers[il].ffn_gate_inp == nullptr) {
+ // feed-forward network (non-MoE)
+ if (model.layers[il].ffn_gate_inp == nullptr) {
- cur = build_norm(ffn_inp,
- model.layers[il].ffn_norm, NULL,
- LLM_NORM_RMS, il);
- cb(cur, "ffn_norm", il);
+ cur = build_norm(ffn_inp,
+ model.layers[il].ffn_norm, NULL,
+ LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
- cur = build_ffn(cur,
- model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
- model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
- model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
- NULL,
- LLM_FFN_SILU, LLM_FFN_PAR, il);
- cb(cur, "ffn_out", il);
+ cur = build_ffn(cur,
+ model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+ model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+ model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+ NULL,
+ LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(cur, "ffn_out", il);
- } else {
- // MoE branch
- cur = build_norm(ffn_inp,
- model.layers[il].ffn_norm, NULL,
- LLM_NORM_RMS, il);
- cb(cur, "ffn_norm", il);
+ } else {
+ // MoE branch
+ cur = build_norm(ffn_inp,
+ model.layers[il].ffn_norm, NULL,
+ LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
- ggml_tensor * moe_out = build_moe_ffn(cur,
- model.layers[il].ffn_gate_inp,
- model.layers[il].ffn_up_exps,
- model.layers[il].ffn_gate_exps,
- model.layers[il].ffn_down_exps,
- nullptr,
- n_expert, n_expert_used,
- LLM_FFN_SILU, true,
- false, 0.0,
- LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
- il);
- cb(moe_out, "ffn_moe_out", il);
+ ggml_tensor * moe_out = build_moe_ffn(cur,
+ model.layers[il].ffn_gate_inp,
+ model.layers[il].ffn_up_exps,
+ model.layers[il].ffn_gate_exps,
+ model.layers[il].ffn_down_exps,
+ nullptr,
+ n_expert, n_expert_used,
+ LLM_FFN_SILU, true,
+ false, 0.0,
+ LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+ il);
+ cb(moe_out, "ffn_moe_out", il);
- // For Granite MoE Shared
- if (hparams.n_ff_shexp > 0) {
- ggml_tensor * ffn_shexp = build_ffn(cur,
- model.layers[il].ffn_up_shexp, NULL, NULL,
- model.layers[il].ffn_gate_shexp, NULL, NULL,
- model.layers[il].ffn_down_shexp, NULL, NULL,
- NULL,
- LLM_FFN_SILU, LLM_FFN_PAR, il);
- cb(ffn_shexp, "ffn_shexp", il);
+ // For Granite MoE Shared
+ if (hparams.n_ff_shexp > 0) {
+ ggml_tensor * ffn_shexp = build_ffn(cur,
+ model.layers[il].ffn_up_shexp, NULL, NULL,
+ model.layers[il].ffn_gate_shexp, NULL, NULL,
+ model.layers[il].ffn_down_shexp, NULL, NULL,
+ NULL,
+ LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(ffn_shexp, "ffn_shexp", il);
- cur = ggml_add(ctx0, moe_out, ffn_shexp);
- cb(cur, "ffn_out", il);
- } else {
- cur = moe_out;
- }
+ cur = ggml_add(ctx0, moe_out, ffn_shexp);
+ cb(cur, "ffn_out", il);
+ } else {
+ cur = moe_out;
}
+ }
- // For Granite architectures - scale residual
+ // For Granite architectures - scale residual
+ if (hparams.f_residual_scale) {
cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
- cur = ggml_add(ctx0, cur, ffn_inp);
- cb(cur, "ffn_out", il);
+ }
+ cur = ggml_add(ctx0, cur, ffn_inp);
+ cb(cur, "ffn_out", il);
- cur = build_cvec(cur, il);
- cb(cur, "l_out", il);
+ cur = build_cvec(cur, il);
+ cb(cur, "l_out", il);
+
+ return cur;
+ }
+};
+
+struct llm_build_granite_hybrid : public llm_graph_context_mamba {
+
+ llm_build_granite_hybrid(
+ const llama_model & model,
+ const llm_graph_params & params,
+ ggml_cgraph * gf) :
+ llm_graph_context_mamba(params) {
+
+ const int64_t n_embd_head = hparams.n_embd_head_v;
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+ ggml_tensor * cur;
+ ggml_tensor * inpL;
+
+ inpL = build_inp_embd(model.tok_embd);
+
+ auto * inp = build_inp_mem_hybrid();
+
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ // Positional embeddings populated if rope enabled
+ ggml_tensor * inp_pos = nullptr;
+ if (hparams.rope_finetuned) {
+ inp_pos = build_inp_pos();
+ }
+
+ for (int il = 0; il < n_layer; ++il) {
+ struct ggml_tensor * inpSA = inpL;
+
+ // norm
+ cur = build_norm(inpL,
+ model.layers[il].attn_norm, NULL,
+ LLM_NORM_RMS, il);
+ cb(cur, "attn_norm", il);
+
+ if (hparams.is_recurrent(il)) {
+ // ssm layer //
+ cur = build_mamba2_layer(inp->get_recr(), gf, cur, model, ubatch, il);
+ } else {
+ // attention layer //
+ cur = build_attention_layer(
+ gf, cur, inp_pos, inp->get_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);
+ }
+
+ // ffn
+ cur = build_layer_ffn(cur, inpSA, model, il);
// input for next layer
inpL = cur;
cur = build_lora_mm(model.output, cur);
// For Granite architectures - scale logits
- cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+ if (hparams.f_logit_scale) {
+ 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 * build_attention_layer(
+ ggml_cgraph * gf,
+ ggml_tensor * cur,
+ ggml_tensor * inp_pos,
+ llm_graph_input_attn_kv_unified * inp_attn,
+ const llama_model & model,
+ const int64_t n_embd_head,
+ const int il) {
+
+ // compute Q and K and (optionally) RoPE them
+ ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+ cb(Qcur, "Qcur", il);
+ if (model.layers[il].bq) {
+ Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+ cb(Qcur, "Qcur", il);
+ }
+
+ ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+ cb(Kcur, "Kcur", il);
+ if (model.layers[il].bk) {
+ Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+ cb(Kcur, "Kcur", il);
+ }
+
+ ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+ cb(Vcur, "Vcur", il);
+ if (model.layers[il].bv) {
+ Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+ cb(Vcur, "Vcur", il);
+ }
+
+ Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
+ Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+ Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+ const bool use_rope = hparams.rope_finetuned;
+ if (use_rope) {
+ 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;
+ cur = build_attn(inp_attn, gf,
+ model.layers[il].wo, model.layers[il].bo,
+ Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+ cb(cur, "attn_out", il);
+ return cur;
+ }
+
+ ggml_tensor * build_layer_ffn(
+ ggml_tensor * cur,
+ ggml_tensor * inpSA,
+ const llama_model & model,
+ const int il) {
+
+ // For Granite architectures - scale residual
+ 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);
+
+ // feed-forward network (non-MoE)
+ if (model.layers[il].ffn_gate_inp == nullptr) {
+
+ cur = build_norm(ffn_inp,
+ model.layers[il].ffn_norm, NULL,
+ LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
+
+ cur = build_ffn(cur,
+ model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+ model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+ model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+ NULL,
+ LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(cur, "ffn_out", il);
+
+ } else {
+ // MoE branch
+ cur = build_norm(ffn_inp,
+ model.layers[il].ffn_norm, NULL,
+ LLM_NORM_RMS, il);
+ cb(cur, "ffn_norm", il);
+
+ ggml_tensor * moe_out = build_moe_ffn(cur,
+ model.layers[il].ffn_gate_inp,
+ model.layers[il].ffn_up_exps,
+ model.layers[il].ffn_gate_exps,
+ model.layers[il].ffn_down_exps,
+ nullptr,
+ n_expert, n_expert_used,
+ LLM_FFN_SILU, true,
+ false, 0.0,
+ LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+ il);
+ cb(moe_out, "ffn_moe_out", il);
+
+ // For Granite MoE Shared
+ if (hparams.n_ff_shexp > 0) {
+ ggml_tensor * ffn_shexp = build_ffn(cur,
+ model.layers[il].ffn_up_shexp, NULL, NULL,
+ model.layers[il].ffn_gate_shexp, NULL, NULL,
+ model.layers[il].ffn_down_shexp, NULL, NULL,
+ NULL,
+ LLM_FFN_SILU, LLM_FFN_PAR, il);
+ cb(ffn_shexp, "ffn_shexp", il);
+
+ cur = ggml_add(ctx0, moe_out, ffn_shexp);
+ cb(cur, "ffn_out", il);
+ } else {
+ cur = moe_out;
+ }
+ }
+
+ // For Granite architectures - scale residual
+ if (hparams.f_residual_scale) {
+ cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+ }
+ cur = ggml_add(ctx0, cur, ffn_inp);
+ cb(cur, "ffn_out", il);
+
+ cur = build_cvec(cur, il);
+ cb(cur, "l_out", il);
+
+ return cur;
+ }
};
// ref: https://github.com/facebookresearch/chameleon
{
llm = std::make_unique<llm_build_granite>(*this, params, gf);
} break;
+ case LLM_ARCH_GRANITE_HYBRID:
+ {
+ llm = std::make_unique<llm_build_granite_hybrid>(*this, params, gf);
+ } break;
case LLM_ARCH_CHAMELEON:
{
llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
case LLM_ARCH_GLM4:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
+ case LLM_ARCH_GRANITE_HYBRID:
case LLM_ARCH_CHAMELEON:
case LLM_ARCH_BAILINGMOE:
case LLM_ARCH_NEO_BERT:
overview / pkg / ggml / sources / llama.cpp / commitdiff