feat: First (broken) pass at end-to-end Bamba implementation

It generates (garbage) tokens! Still lots of debugging to do.

Branch: GraniteFour

Signed-off-by: Gabe Goodhart <[email protected]>

Author: gabe-l-hart

src/llama-model.cpp

@@ -1435,6 +1435,49 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // For Granite MoE Shared
                 ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
             } break;
+        case LLM_ARCH_BAMBA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale, /* required */ false);
+                ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, /* required */ false);
+                ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, /* required */ false);
+                ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, /* required */ false);
+
+                ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+                ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+                ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+                ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+                ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+                // Zero-out n_head_arr and n_head_kv_arr since SSM layers don't
+                // have attention heads. We'll set them correctly below once we
+                // know which layers are attention layers
+                // NOTE: It's important that this happens after n_embd_head_[kv]
+                //  are set above!
+                const auto n_head_attn = hparams.n_head();
+                const auto n_head_kv_attn = hparams.n_head_kv();
+                std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0);
+                std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
+
+                // Attention params
+                std::fill(hparams.recurrent_layer_arr.begin(), hparams.recurrent_layer_arr.end(), true);
+                std::vector<uint32_t> attn_layer_indices;
+                ml.get_arr(LLM_KV_ATTENTION_LAYER_INDICES, attn_layer_indices);
+                for (const auto attn_idx : attn_layer_indices) {
+                    GGML_ASSERT(attn_idx < hparams.n_layer);
+                    hparams.recurrent_layer_arr[attn_idx] = false;
+                    // Correctly set n_head and n_head_kv for attention layers
+                    hparams.n_head_arr[attn_idx] = n_head_attn;
+                    hparams.n_head_kv_arr[attn_idx] = n_head_kv_attn;
+                }
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    // TODO: Add llm type label (not sure this is useful)
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_CHAMELEON:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -3049,6 +3092,83 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_BAMBA:
+                {
+                    // mamba2 Mixer SSM params
+                    // NOTE: int64_t for tensor dimensions
+                    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;
+                    const int64_t n_ssm_head = hparams.ssm_dt_rank;
+                    const int64_t n_group    = hparams.ssm_n_group;
+                    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+
+                    // only an expansion factor of 2 is supported for now
+                    GGML_ASSERT(2 * n_embd == d_inner);
+
+                    // embeddings
+                    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}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+                        if (output == NULL) {
+                            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        // norm
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (hparams.recurrent_layer(i)) {
+                            // ssm layers
+                            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+                            layer.ssm_in_b = create_tensor(tn(LLM_TENSOR_SSM_IN, "bias", i), {n_embd, d_in_proj}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+                            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
+
+                            // no "weight" suffix for these
+                            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
+                            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
+
+                            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
+                            // out_proj
+                            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+                        } else {
+                            // attention layers (with optional bias)
+                            const int64_t n_head_i = hparams.n_head(i);
+                            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
+                            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head_i}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa_i}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa_i}, 0);
+                            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},         llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_k_gqa_i}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_v_gqa_i}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},         llama_model_loader::TENSOR_NOT_REQUIRED);
+                        }
+
+                        // feed forward (w/ optional biases)
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 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}, 0);
+                        layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    }
+                } break;
             case LLM_ARCH_XVERSE:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -12781,6 +12901,160 @@ struct llm_build_granite : public llm_graph_context {
     }
 };
 
+struct llm_build_hybrid_mamba : public llm_graph_context {
+
+    llm_build_hybrid_mamba(
+        const llama_model & model,
+        const llm_graph_params & params,
+        ggml_cgraph * gf,
+        const bool use_mamba2 = true,
+        const bool use_rope = true)
+    : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // Build the inputs in the recurrent cache
+        ggml_tensor * state_copy = build_inp_s_copy();
+
+        // Build the inputs in the attention cache
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        // Positional embeddings populated if rope enabled
+        ggml_tensor * inp_pos = nullptr;
+        if (use_rope) {
+            inp_pos = build_inp_pos();
+        }
+
+        // Extract the recurrent cache from the hybrid parent
+        const auto * kv_recurrent = static_cast<const llama_kv_cache_hybrid *>(memory)->get_child_cache<llama_kv_cache_recurrent>();
+        GGML_ASSERT(kv_recurrent);
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.recurrent_layer(il)) {
+                // ssm layer //
+                if (use_mamba2) {
+                    cur = llm_build_mamba::build_mamba2_layer(this, gf, cur, state_copy, kv_recurrent, model, ubatch, il);
+                } else {
+                    cur = llm_build_mamba::build_mamba_layer(this, gf, cur, state_copy, kv_recurrent, model, ubatch, il);
+                }
+            } else {
+                // attention layer //
+                cur = llm_build_granite::build_attention_layer(
+                    this, gf, cur, inp_pos, inp_attn,
+                    model, n_embd_head, use_rope, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // For Granite architectures - scale residual
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network (non-MoE)
+            if (model.layers[il].ffn_gate_inp == nullptr) {
+
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+            } else {
+                // MoE branch
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                ggml_tensor * moe_out = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, true,
+                        false, 0.0,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                        il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // For Granite MoE Shared
+                if (hparams.n_ff_shexp > 0) {
+                    ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                } else {
+                    cur = moe_out;
+                }
+            }
+
+            // For Granite architectures - scale residual
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        // For Granite architectures - scale logits
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 // ref: https://github.com/facebookresearch/chameleon
 // based on the original build_llama() function, changes:
 //   * qk-norm
@@ -13774,6 +14048,13 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_granite>(*this, params, gf);
             } break;
+        case LLM_ARCH_BAMBA:
+            {
+                llm = std::make_unique<llm_build_hybrid_mamba>(
+                    *this, params, gf,
+                    /* use_mamba2 */ true,
+                    /* use_rope   */ true);
+            } break;
         case LLM_ARCH_CHAMELEON:
             {
                 llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
@@ -13922,6 +14203,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
+        case LLM_ARCH_BAMBA:
         case LLM_ARCH_CHAMELEON:
         case LLM_ARCH_BAILINGMOE:
             return LLAMA_ROPE_TYPE_NORM;