Skip to content

Commit cde8335

Browse files
gabe-l-hartgabe-l-hart
committed
feat: Support GRANITE_MOE_HYBRID in llama-model
This re-uses the Bamba code paths heavily and simply adds the missing parts for loading MoE and the shared expert. Branch: GraniteFour Signed-off-by: Gabe Goodhart <[email protected]>
1 parent acd713a commit cde8335

File tree

1 file changed

+40
-9
lines changed

1 file changed

+40
-9
lines changed

‎src/llama-model.cpp‎

Lines changed: 40 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -1460,6 +1460,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
14601460
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
14611461
} break;
14621462
case LLM_ARCH_BAMBA:
1463+
case LLM_ARCH_GRANITE_MOE_HYBRID:
14631464
{
14641465
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
14651466
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale, /* required */ false);
@@ -1501,6 +1502,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
15011502
// TODO: Add llm type label (not sure this is useful)
15021503
default: type = LLM_TYPE_UNKNOWN;
15031504
}
1505+
1506+
// For Granite MoE Shared
1507+
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
15041508
} break;
15051509
case LLM_ARCH_CHAMELEON:
15061510
{
@@ -3113,6 +3117,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
31133117
}
31143118
} break;
31153119
case LLM_ARCH_BAMBA:
3120+
case LLM_ARCH_GRANITE_MOE_HYBRID:
31163121
{
31173122
// mamba2 Mixer SSM params
31183123
// NOTE: int64_t for tensor dimensions
@@ -3179,14 +3184,31 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
31793184
}
31803185

31813186
// feed forward (w/ optional biases)
3182-
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
3183-
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
3184-
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
3185-
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
3186-
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
3187-
layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
3188-
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
3189-
layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
3187+
if (n_expert > 0) {
3188+
// MoE FFN
3189+
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
3190+
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
3191+
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
3192+
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, TENSOR_NOT_REQUIRED);
3193+
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0);
3194+
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
3195+
3196+
// For Granite MoE Shared
3197+
if (hparams.n_ff_shexp > 0) {
3198+
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
3199+
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
3200+
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
3201+
}
3202+
} else {
3203+
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
3204+
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
3205+
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
3206+
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
3207+
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
3208+
layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
3209+
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
3210+
layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
3211+
}
31903212
}
31913213
} break;
31923214
case LLM_ARCH_XVERSE:
@@ -4635,7 +4657,9 @@ void llama_model::print_info() const {
46354657

46364658
if (arch == LLM_ARCH_MINICPM ||
46374659
arch == LLM_ARCH_GRANITE ||
4638-
arch == LLM_ARCH_GRANITE_MOE) {
4660+
arch == LLM_ARCH_GRANITE_MOE ||
4661+
arch == LLM_ARCH_GRANITE_MOE_HYBRID ||
4662+
arch == LLM_ARCH_BAMBA) {
46394663
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
46404664
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
46414665
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -14035,6 +14059,12 @@ llm_graph_result_ptr llama_model::build_graph(
1403514059
{
1403614060
llm = std::make_unique<llm_build_granite>(*this, params, gf);
1403714061
} break;
14062+
case LLM_ARCH_GRANITE_MOE_HYBRID:
14063+
{
14064+
llm = std::make_unique<llm_build_hybrid_mamba>(*this, params, gf,
14065+
/* use_mamba2 */ true,
14066+
/* use_rope */ false);
14067+
} break;
1403814068
case LLM_ARCH_BAMBA:
1403914069
{
1404014070
llm = std::make_unique<llm_build_hybrid_mamba>(
@@ -14190,6 +14220,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
1419014220
case LLM_ARCH_GLM4:
1419114221
case LLM_ARCH_GRANITE:
1419214222
case LLM_ARCH_GRANITE_MOE:
14223+
case LLM_ARCH_GRANITE_MOE_HYBRID:
1419314224
case LLM_ARCH_BAMBA:
1419414225
case LLM_ARCH_CHAMELEON:
1419514226
case LLM_ARCH_BAILINGMOE:

0 commit comments

Comments
 (0)