1 parent
d1044bc827
commit
9309bf1bcf
20 files changed
+1112
-2
No files matched your search
@@ -102,6 +102,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
||||
{ LLM_ARCH_NEMOTRON, "nemotron" },
|
||||
{ LLM_ARCH_NEMOTRON_H, "nemotron_h" },
|
||||
{ LLM_ARCH_NEMOTRON_H_MOE, "nemotron_h_moe" },
|
||||
{ LLM_ARCH_METIS16, "metis16" },
|
||||
{ LLM_ARCH_EXAONE, "exaone" },
|
||||
{ LLM_ARCH_EXAONE4, "exaone4" },
|
||||
{ LLM_ARCH_EXAONE_MOE, "exaone-moe" },
|
||||
@@ -272,6 +273,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
|
||||
{ LLM_KV_HRM_H_CYCLES, "%s.hrm.h_cycles" },
|
||||
{ LLM_KV_HRM_L_CYCLES, "%s.hrm.l_cycles" },
|
||||
{ LLM_KV_HRM_PREFIX_LM, "%s.hrm.prefix_lm" },
|
||||
{ LLM_KV_METIS_PASS_COUNT, "%s.metis.pass_count" },
|
||||
{ LLM_KV_METIS_MEMORY_LENGTH, "%s.metis.memory_length" },
|
||||
{ LLM_KV_METIS_ROUTE_LENGTH, "%s.metis.route_length" },
|
||||
{ LLM_KV_METIS_CONTINUATION_LOGIT_LIMIT, "%s.metis.continuation_logit_limit" },
|
||||
{ LLM_KV_MUDD_FEED_FORWARD_LENGTH, "%s.mudd.feed_forward_length" },
|
||||
{ LLM_KV_MUDD_TAP_COUNT, "%s.mudd.tap_count" },
|
||||
{ LLM_KV_MUDD_TAP_INDICES, "%s.mudd.tap_indices" },
|
||||
@@ -329,6 +334,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
|
||||
{ LLM_KV_ATTENTION_KV_SOURCE_LAYERS, "%s.attention.kv_source_layers" },
|
||||
{ LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT, "%s.attention.output_group_count" },
|
||||
{ LLM_KV_ATTENTION_OUTPUT_LORA_RANK, "%s.attention.output_lora_rank" },
|
||||
{ LLM_KV_ATTENTION_PASS_LORA_RANK, "%s.attention.pass_lora_rank" },
|
||||
{ LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE, "%s.attention.compress_rope_freq_base" },
|
||||
{ LLM_KV_ATTENTION_COMPRESS_RATIOS, "%s.attention.compress_ratios" },
|
||||
{ LLM_KV_ATTENTION_SHARED_KV_LAYERS, "%s.attention.shared_kv_layers" },
|
||||
@@ -387,6 +393,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
|
||||
{ LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
|
||||
{ LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" },
|
||||
{ LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" },
|
||||
{ LLM_KV_SSM_NORM_EPS, "%s.ssm.norm_epsilon" },
|
||||
|
||||
{ LLM_KV_KDA_HEAD_DIM, "%s.kda.head_dim" },
|
||||
{ LLM_KV_KDA_SAFE_GATE, "%s.kda.safe_gate" },
|
||||
@@ -548,6 +555,36 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
|
||||
{ LLM_TENSOR_FFN_EXP_PROBS_B_VL, "blk.%d.exp_probs_b_vl" },
|
||||
{ LLM_TENSOR_FFN_LATENT_DOWN, "blk.%d.ffn_latent_down" },
|
||||
{ LLM_TENSOR_FFN_LATENT_UP, "blk.%d.ffn_latent_up" },
|
||||
{ LLM_TENSOR_STREAM_EMBD, "stream_embd" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_STATE_WRITE, "depth_mem_state_write" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_META_WRITE, "depth_mem_meta_write" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_ENTRY_BIAS, "depth_mem_entry_bias" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_QUERY, "depth_mem_query" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_KEY, "depth_mem_key" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_VALUE, "depth_mem_value" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_OUT, "depth_mem_out" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_ROUTE_STATE, "depth_mem_route_state" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_ROUTE_RETR, "depth_mem_route_retr" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_ROUTE_DIFF, "depth_mem_route_diff" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_ROUTE_HIST, "depth_mem_route_hist" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_ROUTE, "depth_mem_route" },
|
||||
{ LLM_TENSOR_DEPTH_MEM_STREAM_GATE, "depth_mem_stream_gate" },
|
||||
{ LLM_TENSOR_CONT_HIDDEN_STATE, "cont_hidden_state" },
|
||||
{ LLM_TENSOR_CONT_HIDDEN_RETR, "cont_hidden_retr" },
|
||||
{ LLM_TENSOR_CONT_HIDDEN_DIFF, "cont_hidden_diff" },
|
||||
{ LLM_TENSOR_CONT_HIDDEN_ROUTE, "cont_hidden_route" },
|
||||
{ LLM_TENSOR_CONT_HIDDEN, "cont_hidden" },
|
||||
{ LLM_TENSOR_CONT_OUTPUT, "cont_output" },
|
||||
{ LLM_TENSOR_HC_ATTN_MIX, "blk.%d.hc_attn_mix" },
|
||||
{ LLM_TENSOR_HC_ATTN_READ, "blk.%d.hc_attn_read" },
|
||||
{ LLM_TENSOR_HC_ATTN_WRITE, "blk.%d.hc_attn_write" },
|
||||
{ LLM_TENSOR_HC_FFN_MIX, "blk.%d.hc_ffn_mix" },
|
||||
{ LLM_TENSOR_HC_FFN_READ, "blk.%d.hc_ffn_read" },
|
||||
{ LLM_TENSOR_HC_FFN_WRITE, "blk.%d.hc_ffn_write" },
|
||||
{ LLM_TENSOR_ATTN_PASS_LORA_DOWN, "blk.%d.attn_pass_lora_down" },
|
||||
{ LLM_TENSOR_ATTN_PASS_LORA_UP, "blk.%d.attn_pass_lora_up" },
|
||||
{ LLM_TENSOR_ATTN_PASS_GATE, "blk.%d.attn_pass_gate" },
|
||||
{ LLM_TENSOR_FFN_EXP_EMBD, "blk.%d.ffn_exp_embd" },
|
||||
{ LLM_TENSOR_FFN_PRE_EXP_NORM, "blk.%d.ffn_pre_exp_norm" },
|
||||
{ LLM_TENSOR_FFN_POST_EXP_NORM, "blk.%d.ffn_post_exp_norm" },
|
||||
{ LLM_TENSOR_FFN_RESID_ALPHA, "blk.%d.ffn_resid_alpha" },
|
||||
@@ -1158,6 +1195,36 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
|
||||
// latent projections feed ggml_mul_mat, the buft probe must use MUL_MAT to keep them on GPU
|
||||
{LLM_TENSOR_FFN_LATENT_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_FFN_LATENT_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_STREAM_EMBD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_ADD}},
|
||||
{LLM_TENSOR_DEPTH_MEM_STATE_WRITE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_META_WRITE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_ENTRY_BIAS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_QUERY, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_KEY, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_VALUE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_ROUTE_STATE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_ROUTE_RETR, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_ROUTE_DIFF, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_ROUTE_HIST, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_ROUTE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_DEPTH_MEM_STREAM_GATE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_CONT_HIDDEN_STATE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_CONT_HIDDEN_RETR, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_CONT_HIDDEN_DIFF, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_CONT_HIDDEN_ROUTE, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_CONT_HIDDEN, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_CONT_OUTPUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_HC_ATTN_MIX, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_HC_ATTN_READ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_HC_ATTN_WRITE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_HC_FFN_MIX, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_HC_FFN_READ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_HC_FFN_WRITE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_ATTN_PASS_LORA_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_ATTN_PASS_LORA_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
|
||||
{LLM_TENSOR_ATTN_PASS_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
|
||||
{LLM_TENSOR_FFN_EXP_EMBD, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
|
||||
{LLM_TENSOR_FFN_PRE_EXP_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
|
||||
{LLM_TENSOR_FFN_POST_EXP_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
|
||||
{LLM_TENSOR_FFN_RESID_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
|
||||
@@ -1286,6 +1353,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
|
||||
case LLM_ARCH_LFM3:
|
||||
case LLM_ARCH_NEMOTRON_H:
|
||||
case LLM_ARCH_NEMOTRON_H_MOE:
|
||||
case LLM_ARCH_METIS16:
|
||||
case LLM_ARCH_QWEN3NEXT:
|
||||
case LLM_ARCH_KIMI_LINEAR:
|
||||
case LLM_ARCH_BAILINGMOE3:
|
||||
|
||||
Reference in new issue
Block a user