sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 29 Packages Projects Releases Wiki Activity

llama: end-to-end tests (#19802)

Browse Source 
* tests: add end-to-end tests per model architecture

* fixup for rebase

* fix use-after-free in llama-model-loader.cpp

* fix CI

* fix WebGPU

* fix CI

* disable CI for macOS-latest-cmake-arm64

* use expert_weights_scale only if != 0.0f

* comments
This commit is contained in:
Johannes Gäßler
2026-03-08 12:30:21 +01:00
committed by GitHub
parent a95047979a
commit a976ff081b
33 changed files with 1607 additions and 633 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/baichuan.cpp
+1
View File
@@ -56,6 +56,7 @@ llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_grap
);
break;
case LLM_TYPE_13B:
case LLM_TYPE_UNKNOWN:
break;
default:
GGML_ABORT("fatal error");
src/models/bailingmoe2.cpp
+1 -1
View File
@@ -90,7 +90,7 @@ llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const ll
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);
src/models/dots1.cpp
+1 -1
View File
@@ -91,7 +91,7 @@ llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_para
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);
src/models/exaone-moe.cpp
+1 -1
View File
@@ -100,7 +100,7 @@ llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);
src/models/glm4-moe.cpp
+1 -1
View File
@@ -128,7 +128,7 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(routed_out, "ffn_moe_out", il);
src/models/kimi-linear.cpp
+3 -11
View File
@@ -118,12 +118,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
ggml_build_forward_expand(gf, cur);
// Check layer type by checking which tensors exist
// KDA layers have ssm_a_log tensor, MLA layers have wkv_a_mqa tensor
bool is_kda = (layer.ssm_a != nullptr);
bool is_mla = (layer.wkv_a_mqa != nullptr);
if (is_kda) {
if (hparams.is_recurrent(il)) {
// === KDA Layer (Kimi Delta Attention) with Recurrent State ===
// Reference: vLLM kda.py
const auto * mctx_cur = inp_rs->mctx;
@@ -211,7 +206,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
cur = ggml_mul_mat(ctx0, layer.wo, gated);
cb(cur, "kda_out", il);
} else if (is_mla) {
} else {
// === MLA Layer (Multi-head Latent Attention) without KV Cache ===
// Reference: vLLM mla.py
// Step 1: Q projection and reshape
@@ -310,9 +305,6 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
cur = build_attn(inp_attn_kv, layer.wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale_mla, il);
cb(cur, "mla_out", il);
}
} else {
// Unknown layer type - this should not happen
GGML_ABORT("Kimi layer is neither KDA nor MLA - missing required tensors");
}
// On last layer, select only the output tokens
@@ -349,7 +341,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
hparams.n_expert,
hparams.n_expert_used,
LLM_FFN_SILU, true,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);
src/models/mamba-base.cpp
+4
View File
@@ -30,6 +30,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba_layer(llm_graph_input_rs * inp,
GGML_ASSERT(n_seqs != 0);
GGML_ASSERT(ubatch.equal_seqs());
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
GGML_ASSERT(d_inner % n_head == 0);
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
@@ -154,6 +155,7 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
const auto kv_head = mctx_cur->get_head();
const int64_t n_embd = hparams.n_embd;
const int64_t d_conv = hparams.ssm_d_conv;
const int64_t d_inner = hparams.ssm_d_inner;
const int64_t d_state = hparams.ssm_d_state;
@@ -167,6 +169,8 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
GGML_ASSERT(n_seqs != 0);
GGML_ASSERT(ubatch.equal_seqs());
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
GGML_ASSERT(d_inner % n_head == 0);
GGML_ASSERT(d_inner % (n_group*n_embd) == 0);
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
src/models/nemotron-h.cpp
+1 -1
View File
@@ -124,7 +124,7 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
hparams.expert_weights_scale, hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
il);
cb(moe_out, "ffn_moe_out", il);
src/models/plamo2.cpp
+3 -1
View File
@@ -27,7 +27,7 @@ llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_pa
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
// check if this layer is Mamba or Attention
bool is_mamba_layer = hparams.is_recurrent(il);
const bool is_mamba_layer = hparams.is_recurrent(il);
if (is_mamba_layer) {
// PLaMo-2 Mamba layer
@@ -171,6 +171,8 @@ ggml_tensor * llm_build_plamo2::build_plamo2_mamba_layer(llm_graph_input_rs * in
GGML_ASSERT(n_seqs != 0);
GGML_ASSERT(ubatch.equal_seqs());
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
GGML_ASSERT(d_inner % n_head == 0);
GGML_ASSERT(n_group == 0);
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);