mtmd, llama: add GLM4V vision-language model support

eelbaz · claude · eelbaz · commit ccf0bb52a99e · 2025-12-12T22:25:33.000-05:00
Add complete GLM4V (GLM-4.6V-Flash) support including: **Vision Encoder (mtmd):** - Dual Conv2D patch embedding (simulating Conv3D temporal reduction) - M-RoPE using ggml_rope_multi() with [h,w,h,w] position pattern - 2x2 patch merger with downsample convolution - SwiGLU-based merger FFN **LLM Architecture (libllama):** - New LLM_ARCH_GLM4V based on GLM4 with M-RoPE - Uses LLAMA_ROPE_TYPE_MROPE with rope_sections from model config - Reuses ggml_rope_multi() (same as Qwen2VL) for position encoding Key design decisions: - Vision encoder uses ggml_rope_multi() instead of custom RoPE - LLM follows GLM4 structure with M-RoPE (not Qwen2VL structure) - Minimal code: ~300 lines total across all files Tested with GLM-4.6V-Flash producing correct image descriptions. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -71,6 +71,7 @@ add_library(llama
             models/gemma3n-iswa.cpp
             models/glm4-moe.cpp
             models/glm4.cpp
+            models/glm4v.cpp
             models/gpt2.cpp
             models/gptneox.cpp
             models/granite-hybrid.cpp
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
@@ -69,6 +69,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_CHATGLM,          "chatglm"          },
     { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_GLM4_MOE,         "glm4moe"          },
+    { LLM_ARCH_GLM4V,            "glm4v"            },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -1634,6 +1635,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
         },
     },
+    {
+        LLM_ARCH_GLM4V,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
diff --git a/src/llama-arch.h b/src/llama-arch.h
@@ -73,6 +73,7 @@ enum llm_arch {
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
     LLM_ARCH_GLM4_MOE,
+    LLM_ARCH_GLM4V,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
@@ -1723,6 +1723,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GLM4V:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
+                switch (hparams.n_layer) {
+                    case 40: type = LLM_TYPE_9B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_BITNET:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -5142,6 +5151,45 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 }
                 break;
+            case LLM_ARCH_GLM4V:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        }
+
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, 0);
+
+                        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_NEMOTRON:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -7423,6 +7471,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_glm4_moe>(*this, params);
             } break;
+        case LLM_ARCH_GLM4V:
+            {
+                llm = std::make_unique<llm_build_glm4v>(*this, params);
+            } break;
         case LLM_ARCH_BITNET:
             {
                 llm = std::make_unique<llm_build_bitnet>(*this, params);
@@ -7832,6 +7884,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:
+        case LLM_ARCH_GLM4V:
             return LLAMA_ROPE_TYPE_MROPE;
         case LLM_ARCH_QWEN3VL:
         case LLM_ARCH_QWEN3VLMOE:
diff --git a/src/models/glm4v.cpp b/src/models/glm4v.cpp
@@ -0,0 +1,133 @@
+#include "models.h"
+
+llm_build_glm4v::llm_build_glm4v(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    // M-RoPE sections from hparams
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            }
+
+            // GLM4V uses M-RoPE (multi-dimensional rotary position embeddings)
+            Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
+                                   n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                   ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
+                                   n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                   ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // Post-attention norm
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Add the input (residual connection after post-attention norm)
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            // Pre-MLP norm
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // MLP
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            // Post-MLP norm
+            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_mlp_norm", il);
+        }
+        // Add residual connection after post-MLP norm
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+    // Final norm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // Output projection
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
diff --git a/src/models/models.h b/src/models/models.h
@@ -222,6 +222,10 @@ struct llm_build_glm4_moe : public llm_graph_context {
     llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_glm4v : public llm_graph_context {
+    llm_build_glm4v(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_gpt2 : public llm_graph_context {
     llm_build_gpt2(const llama_model & model, const llm_graph_params & params);
 };
diff --git a/tools/mtmd/CMakeLists.txt b/tools/mtmd/CMakeLists.txt
@@ -15,6 +15,7 @@ add_library(mtmd
             clip-graph.h
             models/models.h
             models/cogvlm.cpp
+            models/glm4v.cpp
             models/internvl.cpp
             models/kimivl.cpp
             models/llama4.cpp
diff --git a/tools/mtmd/clip-impl.h b/tools/mtmd/clip-impl.h
@@ -133,6 +133,15 @@
 #define TN_TOK_BOI         "v.boi"
 #define TN_TOK_EOI         "v.eoi"
 
+// glm4v
+#define TN_GLM4V_POST_CONV_LN   "mm.post_conv_ln.%s"
+#define TN_GLM4V_DOWNSAMPLE     "mm.downsample.%s"
+#define TN_GLM4V_MERGER_PROJ    "mm.merger.proj.%s"
+#define TN_GLM4V_MERGER_NORM    "mm.merger.norm.%s"
+#define TN_GLM4V_MERGER_GATE    "mm.merger.gate.%s"
+#define TN_GLM4V_MERGER_UP      "mm.merger.up.%s"
+#define TN_GLM4V_MERGER_DOWN    "mm.merger.down.%s"
+
 // align x to upper multiple of n
 #define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
 
@@ -164,6 +173,7 @@ enum projector_type {
     PROJECTOR_TYPE_LIGHTONOCR,
     PROJECTOR_TYPE_COGVLM,
     PROJECTOR_TYPE_JANUS_PRO,
+    PROJECTOR_TYPE_GLM4V,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -190,6 +200,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
     { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
     { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
+    { PROJECTOR_TYPE_GLM4V,     "glm4v"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {
diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.h
@@ -267,6 +267,22 @@ struct clip_model {
     ggml_tensor * mm_boi = nullptr;
     ggml_tensor * mm_eoi = nullptr;
 
+    // glm4v
+    ggml_tensor * mm_post_conv_ln_w = nullptr;
+    ggml_tensor * mm_post_conv_ln_b = nullptr;
+    ggml_tensor * mm_downsample_w = nullptr;
+    ggml_tensor * mm_downsample_b = nullptr;
+    ggml_tensor * mm_merger_proj_w = nullptr;
+    ggml_tensor * mm_merger_proj_b = nullptr;
+    ggml_tensor * mm_merger_norm_w = nullptr;
+    ggml_tensor * mm_merger_norm_b = nullptr;
+    ggml_tensor * mm_merger_gate_w = nullptr;
+    ggml_tensor * mm_merger_gate_b = nullptr;
+    ggml_tensor * mm_merger_up_w = nullptr;
+    ggml_tensor * mm_merger_up_b = nullptr;
+    ggml_tensor * mm_merger_down_w = nullptr;
+    ggml_tensor * mm_merger_down_b = nullptr;
+
     bool audio_has_avgpool() const {
         return proj_type == PROJECTOR_TYPE_QWEN2A
             || proj_type == PROJECTOR_TYPE_VOXTRAL;
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
diff --git a/tools/mtmd/models/glm4v.cpp b/tools/mtmd/models/glm4v.cpp
diff --git a/tools/mtmd/models/models.h b/tools/mtmd/models/models.h
