Skip to content

GLM-5.2 quantization crash and sparse DSA indexer loading #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
csabakecskemeti wants to merge 3 commits into
base: master
Choose a base branch
from
+17 −10
Open

GLM-5.2 quantization crash and sparse DSA indexer loading #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
22ea8d8
llama-quant : fix FFN layer bounds crash for models with nextn/MTP bl…
csabakecskemeti Jun 18, 2026
ed6dcaf
glm-dsa : fix missing indexer tensors on sparse-indexer blocks
csabakecskemeti Jun 18, 2026
40e9bb4
convert : bump transformers requirement to >=5.0.0
csabakecskemeti Jun 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion requirements/requirements-convert_legacy_llama.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,9 @@
numpy~=1.26.4
sentencepiece>=0.1.98,<0.3.0

transformers==4.57.6
# GLM-5.2 and other recent models require transformers 5.x; the TokenizersBackend class
# referenced in their tokenizer_config.json does not exist in 4.x
transformers>=5.0.0

gguf>=0.1.0
protobuf>=4.21.0,<5.0.0
6 changes: 5 additions & 1 deletion src/llama-quant.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -847,7 +847,11 @@ static void init_quantize_state_counters(quantize_state_impl & qs, std::vector<t
qs.has_tied_embeddings = false;
}
}
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer();
// Use n_layer_all, not n_layer(), because models with nextn/MTP blocks (e.g. GLM-5.2,
// BailingMoE2) store ffn_down/gate/up_exps tensors for those extra blocks in the GGUF
// file. The MoE path in layer_info() parses the block index from the tensor name, so
// those tensors would exceed the bounds set by n_layer() = n_layer_all - n_layer_nextn.
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)qs.model.hparams.n_layer_all;
}

//
Expand Down
4 changes: 2 additions & 2 deletions src/models/deepseek32.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,8 +219,8 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_

ggml_tensor * top_k = nullptr;

// lightning indexer
{
// lightning indexer - skipped when tensors are absent (e.g. GLM-5.2 only has them on certain blocks)
if (model.layers[il].indexer_attn_q_b) {
ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
cb(indexer_q, "indexer_q", il);

Expand Down
13 changes: 7 additions & 6 deletions src/models/glm-dsa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,12 +100,13 @@ void llama_model_glm_dsa::load_arch_tensors(llama_model_loader &) {

layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);

// DSA indexer
layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {hparams.indexer_head_size}, flags);
layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {hparams.indexer_head_size}, flags);
layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, hparams.indexer_n_head}, flags);
layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, hparams.indexer_head_size}, flags);
layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
// DSA indexer - only present on a subset of blocks (dense-lead blocks + every 4th MoE block),
// so mark all as TENSOR_NOT_REQUIRED; the graph skips the indexer path when these are null.
layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, hparams.indexer_n_head}, flags | TENSOR_NOT_REQUIRED);
layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
if (i < (int) hparams.n_layer_dense_lead) {
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags);
Expand Down
Loading