LLM inference in C/C++
C++ ggml
model: fix all granite family parameter counts (#28643)
Aaron Teod7e86430a7d5
diff --git a/src/llama-model.cpp b/src/llama-model.cppindex 54009b3696..d10b60afd9 100644--- a/src/llama-model.cpp+++ b/src/llama-model.cpp@@ -936,6 +936,7 @@ const char * llm_type_name(llm_type type) {case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)";case LLM_TYPE_A13B: return "A13B";case LLM_TYPE_1B_A400M: return "1B.A400M";+ case LLM_TYPE_3B_A800M: return "3B.A800M";case LLM_TYPE_7B_A1B: return "7B.A1B";case LLM_TYPE_8B_A1B: return "8B.A1B";case LLM_TYPE_7_9B_A1_3B: return "7.9B.A1.3B";@@ -946,6 +947,7 @@ const char * llm_type_name(llm_type type) {case LLM_TYPE_26B_A4B: return "26B.A4B";case LLM_TYPE_30B_A3B: return "30B.A3B";case LLM_TYPE_31B_A3_5B: return "31B.A3.5B";+ case LLM_TYPE_32B_A9B: return "32B.A9B";case LLM_TYPE_35B_A3B: return "35B.A3B";case LLM_TYPE_48B_A3B: return "48B.A3B";case LLM_TYPE_75B_A9B: return "75B.A9B";diff --git a/src/llama-model.h b/src/llama-model.hindex c0cc406556..a02b30ca7a 100644--- a/src/llama-model.h+++ b/src/llama-model.h@@ -117,6 +117,7 @@ enum llm_type {LLM_TYPE_17B_128E, // llama4 MaverickLLM_TYPE_A13B,LLM_TYPE_1B_A400M, // Granite3 MoE+ LLM_TYPE_3B_A800M, // Granite3 MoELLM_TYPE_7B_A1B,LLM_TYPE_8B_A1B, // lfm2moeLLM_TYPE_7_9B_A1_3B, // Ling-3.0-tiny@@ -127,6 +128,7 @@ enum llm_type {LLM_TYPE_26B_A4B, // Gemma4LLM_TYPE_30B_A3B,LLM_TYPE_31B_A3_5B,+ LLM_TYPE_32B_A9B, // Granite4 HybridLLM_TYPE_35B_A3B, // Qwen3.5LLM_TYPE_48B_A3B, // Kimi LinearLLM_TYPE_75B_A9B, // Nemotron 3 Puzzlediff --git a/src/models/granite-hybrid.cpp b/src/models/granite-hybrid.cppindex 8a8f7e19ff..c177ae7875 100644--- a/src/models/granite-hybrid.cpp+++ b/src/models/granite-hybrid.cpp@@ -30,7 +30,7 @@ void llama_model_granite_hybrid::load_arch_hparams(llama_model_loader & ml) {case 768: type = LLM_TYPE_350M; break;case 1536: type = (hparams.n_ff() == 512 ? LLM_TYPE_7B_A1B : LLM_TYPE_1B); break;case 2048: case 2560: type = LLM_TYPE_3B; break;- case 4096: type = LLM_TYPE_32B; break;+ case 4096: type = LLM_TYPE_32B_A9B; break;default: type = LLM_TYPE_UNKNOWN;}diff --git a/src/models/granite-moe.cpp b/src/models/granite-moe.cppindex 156553edfd..febe1bfa7a 100644--- a/src/models/granite-moe.cpp+++ b/src/models/granite-moe.cpp@@ -9,8 +9,7 @@ void llama_model_granite_moe::load_arch_hparams(llama_model_loader & ml) {switch (hparams.n_layer()) {case 24: type = LLM_TYPE_1B_A400M; break;- case 32: type = LLM_TYPE_3B; break;- case 40: type = LLM_TYPE_3B; break;+ case 32: type = LLM_TYPE_3B_A800M; break;// Add additional layer/vocab/etc checks here for other model sizesdefault: type = LLM_TYPE_UNKNOWN;}diff --git a/src/models/granite.cpp b/src/models/granite.cppindex 9e9f97e94d..60d463aeda 100644--- a/src/models/granite.cpp+++ b/src/models/granite.cpp@@ -38,7 +38,16 @@ void llama_model_granite::load_arch_hparams(llama_model_loader & ml) {switch (hparams.n_layer()) {case 32: type = LLM_TYPE_3B; break;- case 40: type = LLM_TYPE_3B; break;+ case 40: {+ switch (hparams.n_embd) {+ case 2048: type = LLM_TYPE_2B; break;+ case 2560: type = LLM_TYPE_3B; break;+ case 4096: type = LLM_TYPE_8B; break;+ default: type = LLM_TYPE_UNKNOWN;+ }+ break;+ }+ case 64: type = LLM_TYPE_30B; break;// Add additional layer/vocab/etc checks here for other model sizesdefault: type = LLM_TYPE_UNKNOWN;}