Skip to content

model : support NVFP4 tensors for Gemma4 #21971

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

Merged
CISC merged 4 commits into from
Apr 16, 2026
Merged

model : support NVFP4 tensors for Gemma4 #21971

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0a40e49
support nvfp4 tensors for Gemma4
CISC Apr 15, 2026
75f3e23
add wo_s to build_attn
CISC Apr 15, 2026
69a2478
add wo_s to build_attn
CISC Apr 15, 2026
8ece3eb
fix glm4
CISC Apr 15, 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
25 changes: 20 additions & 5 deletions src/llama-graph.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2011,6 +2011,7 @@ ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_no_cache * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
Expand Down Expand Up @@ -2044,7 +2045,7 @@ ggml_tensor * llm_graph_context::build_attn(
cb(cur, "kqv_out", il);

if (wo) {
cur = build_lora_mm(wo, cur);
cur = build_lora_mm(wo, cur, wo_s);
}

if (wo_b) {
Expand Down Expand Up @@ -2095,6 +2096,7 @@ ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_kv * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
Expand Down Expand Up @@ -2146,10 +2148,15 @@ ggml_tensor * llm_graph_context::build_attn(
}

if (wo) {
cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
// GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
cur = build_lora_mm(wo, cur);
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
if (wo_s) {
cur = ggml_mul(ctx0, cur, wo_s);
}
} else {
cur = build_lora_mm(wo, cur, wo_s);
}
}

Expand Down Expand Up @@ -2193,6 +2200,7 @@ ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_k * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
Expand Down Expand Up @@ -2227,10 +2235,15 @@ ggml_tensor * llm_graph_context::build_attn(
cb(cur, "kqv_out", il);

if (wo) {
cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
cur = build_lora_mm(wo, cur);
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
if (wo_s) {
cur = ggml_mul(ctx0, cur, wo_s);
}
} else {
cur = build_lora_mm(wo, cur, wo_s);
Comment on lines 2238 to +2246
Copy link
Copy Markdown
Member

@ggerganov ggerganov Apr 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe a follow-up PR to fix the order of the build_lora_mm arguments (e.g. cur, wo, wo_s) and add an optional precision argument to avoid this branching.

Copy link
Copy Markdown
Member Author

@CISC CISC Apr 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, will be more manageable after merging #21245

}
}

Expand All @@ -2245,6 +2258,7 @@ ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_kv_iswa * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
Expand Down Expand Up @@ -2313,7 +2327,7 @@ ggml_tensor * llm_graph_context::build_attn(
}

if (wo) {
cur = build_lora_mm(wo, cur);
cur = build_lora_mm(wo, cur, wo_s);
}

if (wo_b) {
Expand Down Expand Up @@ -2344,6 +2358,7 @@ ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_cross * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
Expand All @@ -2368,7 +2383,7 @@ ggml_tensor * llm_graph_context::build_attn(
cb(cur, "kqv_out", il);

if (wo) {
cur = build_lora_mm(wo, cur);
cur = build_lora_mm(wo, cur, wo_s);
}

if (wo_b) {
Expand Down
5 changes: 5 additions & 0 deletions src/llama-graph.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -892,6 +892,7 @@ struct llm_graph_context {
llm_graph_input_attn_no_cache * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
Expand All @@ -907,6 +908,7 @@ struct llm_graph_context {
llm_graph_input_attn_kv * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
Expand All @@ -922,6 +924,7 @@ struct llm_graph_context {
llm_graph_input_attn_k * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
Expand All @@ -938,6 +941,7 @@ struct llm_graph_context {
llm_graph_input_attn_kv_iswa * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
Expand All @@ -953,6 +957,7 @@ struct llm_graph_context {
llm_graph_input_attn_cross * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * wo_s,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
Expand Down
4 changes: 2 additions & 2 deletions src/models/afmoe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,7 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);

cur = build_attn(inp_attn,
NULL, NULL, // wo will be applied after gating
NULL, NULL, NULL, // wo will be applied after gating
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);

Expand All @@ -91,7 +91,7 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
cb(cur, "attn_gated", il);

// now apply output projection
cur = build_lora_mm(model.layers[il].wo, cur);
cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
cb(cur, "attn_o_proj", il);
}

Expand Down
4 changes: 1 addition & 3 deletions src/models/apertus.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,5 @@
#include "models.h"



llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();

Expand Down Expand Up @@ -62,7 +60,7 @@ llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur_pos", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
Expand Down
3 changes: 1 addition & 2 deletions src/models/arcee.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,5 @@
#include "models.h"


llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();

Expand Down Expand Up @@ -78,7 +77,7 @@ llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
Expand Down
2 changes: 1 addition & 1 deletion src/models/arctic.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,7 @@ llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_pa
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
3 changes: 1 addition & 2 deletions src/models/baichuan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,5 @@
#include "models.h"


llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();

Expand Down Expand Up @@ -67,7 +66,7 @@ llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/bailingmoe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,7 +70,7 @@ llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/bailingmoe2.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,7 @@ llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const ll
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/bert.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,7 +100,7 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
cb(cur, "kqv_out", il);
}
Expand Down
2 changes: 1 addition & 1 deletion src/models/bitnet.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,7 +73,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
NULL, NULL,
NULL, NULL, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);

cur = build_norm(cur,
Expand Down
2 changes: 1 addition & 1 deletion src/models/bloom.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,7 @@ llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/chameleon.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, nullptr,
model.layers[il].wo, nullptr, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/chatglm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,7 @@ llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/codeshell.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,7 @@ llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
6 changes: 4 additions & 2 deletions src/models/cogvlm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,18 +28,20 @@ llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_pa

for (int il = 0; il < n_layer; ++il) {
// get either the text or image weight tensors
ggml_tensor *wqkv, *wo;
ggml_tensor *wqkv, *wo, *wo_s;
ggml_tensor *ffn_gate, *ffn_down, *ffn_up;

if (is_text) {
wqkv = model.layers[il].wqkv;
wo = model.layers[il].wo;
wo_s = model.layers[il].wo_s;
ffn_gate = model.layers[il].ffn_gate;
ffn_down = model.layers[il].ffn_down;
ffn_up = model.layers[il].ffn_up;
} else {
wqkv = model.layers[il].visexp_attn_wqkv;
wo = model.layers[il].visexp_attn_wo;
wo_s = nullptr;
ffn_gate = model.layers[il].visexp_ffn_gate;
ffn_down = model.layers[il].visexp_ffn_down;
ffn_up = model.layers[il].visexp_ffn_up;
Expand All @@ -64,7 +66,7 @@ llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_pa
Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);

cur = build_attn(inp_attn,
wo, nullptr,
wo, nullptr, wo_s,
Qcur, Kcur, Vcur,
nullptr, nullptr, nullptr,
kq_scale, il);
Expand Down
2 changes: 1 addition & 1 deletion src/models/cohere2-iswa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,7 @@ llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
2 changes: 1 addition & 1 deletion src/models/command-r.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,7 +73,7 @@ llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {
Expand Down
2 changes: 1 addition & 1 deletion src/models/dbrx.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

Expand Down
4 changes: 1 addition & 3 deletions src/models/deci.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,5 @@
#include "models.h"



llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();

Expand Down Expand Up @@ -80,7 +78,7 @@ llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1 && inp_out_ids) {
Expand Down
2 changes: 1 addition & 1 deletion src/models/deepseek.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,7 +68,7 @@ llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);

cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1 && inp_out_ids) {
Expand Down
6 changes: 3 additions & 3 deletions src/models/deepseek2.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,7 +84,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
cb(Kcur, "k_pe", il);

cur = build_attn(inp_attn_kv,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}
Expand Down Expand Up @@ -182,7 +182,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr

// note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
cur = build_attn(inp_attn_k,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
} else {
ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
Expand Down Expand Up @@ -219,7 +219,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr

// note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
cur = build_attn(inp_attn_kv,
model.layers[il].wo, NULL,
model.layers[il].wo, NULL, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
}
Expand Down
Loading