From 6dcc21e7f5a7017ff1c4c54cb709d3bca16e0ff5 Mon Sep 17 00:00:00 2001 From: Molly Sophia Date: Wed, 15 Jan 2025 20:43:23 +0800 Subject: [PATCH] WIP: Add support for rwkv v7 Signed-off-by: Molly Sophia --- convert_hf_to_gguf.py | 75 +++++++++++ ggml/include/ggml.h | 11 ++ ggml/src/ggml-cpu/ggml-cpu.c | 185 ++++++++++++++++++++++++- ggml/src/ggml.c | 54 +++++++- gguf-py/gguf/constants.py | 93 ++++++++++--- gguf-py/gguf/gguf_writer.py | 12 ++ gguf-py/gguf/tensor_mapping.py | 89 ++++++++++-- src/llama-arch.cpp | 83 +++++++++--- src/llama-arch.h | 17 +++ src/llama-hparams.h | 4 + src/llama-model.cpp | 89 ++++++++++++ src/llama-model.h | 14 ++ src/llama.cpp | 238 +++++++++++++++++++++++++++++++++ tests/test-backend-ops.cpp | 36 +++++ 14 files changed, 952 insertions(+), 48 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 018a2a588..c427263d7 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -3480,6 +3480,81 @@ class RWKV6Qwen2Model(Rwkv6Model): yield (new_name, data) +@Model.register("Rwkv7ForCausalLM") +class Rwkv7Model(Rwkv6Model): + model_arch = gguf.MODEL_ARCH.RWKV7 + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + head_size = self.hparams["head_size"] + hidden_size = self.hparams["hidden_size"] + layer_norm_eps = self.hparams["layer_norm_epsilon"] + intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else (hidden_size * 4) + + # ICLR: In-Context-Learning-Rate + calc_lora_rank = lambda exponent, multiplier: max(1, round(hidden_size ** exponent * multiplier / 32)) * 32 + lora_rank_decay = self.hparams["lora_rank_decay"] if self.hparams["lora_rank_decay"] is not None else calc_lora_rank(0.5, 1.8) + lora_rank_iclr = self.hparams["lora_rank_iclr"] if self.hparams["lora_rank_iclr"] is not None else calc_lora_rank(0.5, 1.8) + lora_rank_value_residual_mix = self.hparams["lora_rank_value_residual_mix"] if self.hparams["lora_rank_value_residual_mix"] is not None else calc_lora_rank(0.5, 1.3) + lora_rank_gate = self.hparams["lora_rank_gate"] if self.hparams["lora_rank_gate"] is not None else calc_lora_rank(0.8, 0.6) + + # RWKV isn't context limited + self.gguf_writer.add_context_length(1048576) + self.gguf_writer.add_embedding_length(hidden_size) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_layer_norm_eps(layer_norm_eps) + self.gguf_writer.add_wkv_head_size(head_size) + self.gguf_writer.add_decay_lora_rank(lora_rank_decay) + self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr) + self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix) + self.gguf_writer.add_gate_lora_rank(lora_rank_gate) + self.gguf_writer.add_feed_forward_length(intermediate_size) + self.gguf_writer.add_file_type(self.ftype) + + # required by llama.cpp, unused + self.gguf_writer.add_head_count(0) + + lerp_weights: dict[int, dict[str, Tensor]] = {} + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if bid is not None and "attention.x_" in name: + try: + self.lerp_weights[bid][name] = data_torch + except KeyError: + self.lerp_weights[bid] = {name: data_torch} + if all(f"model.blocks.{bid}.attention.x_{i}" in self.lerp_weights[bid].keys() for i in ["r", "w", "k", "v", "a", "g"]): + new_name = f"blk.{bid}.time_mix_lerp_fused.weight" + data = torch.stack([self.lerp_weights[bid][f"model.blocks.{bid}.attention.x_{i}"].squeeze(0) for i in ["r", "w", "k", "v", "a", "g"]], dim=0) + yield (new_name, data) + return + else: + data_torch = data_torch.squeeze() + new_name = self.map_tensor_name(name) + + if not (new_name.endswith(".weight") or new_name.endswith(".bias")): + new_name += ".weight" + + if any( + new_name.endswith(t) for t in [ + "time_mix_w1.weight", "time_mix_w2.weight", + "time_mix_a1.weight", "time_mix_a2.weight", + "time_mix_v1.weight", "time_mix_v2.weight", + "time_mix_g1.weight", "time_mix_g2.weight", + ] + ): + data_torch = data_torch.transpose(0, 1) + + if 'r_k' in new_name: + data_torch = data_torch.flatten() + + if bid == 0 and "time_mix_a" in new_name: + # dummy v0/v1/v2 on first layer + # easist way to make llama happy + yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch) + + yield (new_name, data_torch) + + @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM") class MambaModel(Model): model_arch = gguf.MODEL_ARCH.MAMBA diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index f42dd1c00..c58494d1c 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -502,6 +502,7 @@ extern "C" { GGML_OP_GET_REL_POS, GGML_OP_ADD_REL_POS, GGML_OP_RWKV_WKV6, + GGML_OP_RWKV_WKV7, GGML_OP_GATED_LINEAR_ATTN, GGML_OP_UNARY, @@ -1894,6 +1895,16 @@ extern "C" { struct ggml_tensor * td, struct ggml_tensor * state); + GGML_API struct ggml_tensor * ggml_rwkv_wkv7( + struct ggml_context * ctx, + struct ggml_tensor * r, + struct ggml_tensor * w, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * state); + GGML_API struct ggml_tensor * ggml_gated_linear_attn( struct ggml_context * ctx, struct ggml_tensor * k, diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 356502cb3..1b6127b8e 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -12129,6 +12129,184 @@ static void ggml_compute_forward_rwkv_wkv6( } } +// ggml_compute_forward_rwkv_wkv7 + +static void ggml_compute_forward_rwkv_wkv7_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + const int64_t T = dst->src[1]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t HEADS = dst->src[1]->ne[1]; + const int64_t n_seqs = dst->src[6]->ne[1]; + const int64_t head_size = C / HEADS; + + float * dst_data = (float *) dst->data; + float * state = ((float *) dst->data) + C * T; + + const int ith = params->ith; + const int nth = params->nth; + + if (ith >= HEADS) { + return; + } + + const int h_start = (HEADS * ith) / nth; + const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ? + (HEADS * (ith + 1)) / nth : HEADS; + + float * r = (float *) dst->src[0]->data; + float * w = (float *) dst->src[1]->data; + float * k = (float *) dst->src[2]->data; + float * v = (float *) dst->src[3]->data; + float * a = (float *) dst->src[4]->data; + float * b = (float *) dst->src[5]->data; + + int64_t t_stride = HEADS * head_size; // Same to C + + int64_t h_stride = C / HEADS; + GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS + int64_t h_stride_2d = head_size * head_size; + + #if defined(GGML_SIMD) + for (int64_t t = 0; t < T; t++) { + int64_t t_offset = t * t_stride; + int64_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + int64_t h_offset = h * h_stride; + int64_t t_h_offset = t_offset + h_offset; + int64_t h_2d_offset = h * h_stride_2d; + + for (int64_t ii = 0; ii < head_size; ii++) { + int64_t t_h_i_offset = t_h_offset + ii; + int64_t h_2d_i_offset = h_2d_offset + ii * h_stride; + + GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]); + + float sa = 0; + { + GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO }; + GGML_F32_VEC ax[GGML_F32_ARR]; + GGML_F32_VEC ay[GGML_F32_ARR]; + for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) { + for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) { + ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]); + ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]); + sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]); + } + } + GGML_F32_VEC_REDUCE(sa, sum); + } + + GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa); + + int64_t j = 0; + GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO }; + for (; j < head_size; j += GGML_F32_STEP) { + for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) { + int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR; + int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR; + + GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]); + GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]); + GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]); + GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]); + + k_vec = GGML_F32_VEC_MUL(v_vec, k_vec); + + GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]); + // kv + s * decay + sa * b + state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec); + state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec); + GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec); + + result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec); + } + } + GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec); + + // There shouldn't be left-overs though. + for (; j < head_size; j++) { + int64_t t_h_j_offset = t_h_offset + j; + int64_t h_2d_i_j_offset = h_2d_i_offset + j; + + float r_val = r[t_h_j_offset]; + float w_val = w[t_h_j_offset]; + float k_val = k[t_h_j_offset]; + float b_val = b[t_h_j_offset]; + float kv_val = v[t_h_i_offset] * k_val; + + float prev_state_val = state_prev[h_2d_i_j_offset]; + state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val; + dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val; + } + } + } + } + #else + for (int64_t t = 0; t < T; t++) { + int64_t t_offset = t * t_stride; + int64_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + int64_t h_offset = h * h_stride; + int64_t t_h_offset = t_offset + h_offset; + int64_t h_2d_offset = h * h_stride_2d; + + for (int64_t i = 0; i < head_size; i++) { + int64_t t_h_i_offset = t_h_offset + i; + int64_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float v_val = v[t_h_i_offset]; + + float sa = 0, result = 0; + for (int64_t j = 0; j < head_size; j++) { + sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j]; + } + + for (int64_t j = 0; j < head_size; j++) { + int64_t t_h_j_offset = t_h_offset + j; + int64_t h_2d_i_j_offset = h_2d_i_offset + j; + + float r_val = r[t_h_j_offset]; + float w_val = w[t_h_j_offset]; + float k_val = k[t_h_j_offset]; + float b_val = b[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val; + result += state_cur[h_2d_i_j_offset] * r_val; + } + dst_data[t_h_i_offset] = result; + } + } + } + #endif +} + + +static void ggml_compute_forward_rwkv_wkv7( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rwkv_wkv7_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + // ggml_compute_forward_gla static void ggml_compute_forward_gla_f32( @@ -13073,6 +13251,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_rwkv_wkv6(params, tensor); } break; + case GGML_OP_RWKV_WKV7: + { + ggml_compute_forward_rwkv_wkv7(params, tensor); + } break; case GGML_OP_GATED_LINEAR_ATTN: { ggml_compute_forward_gla(params, tensor); @@ -13369,13 +13551,14 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_FLASH_ATTN_BACK: case GGML_OP_SSM_CONV: case GGML_OP_SSM_SCAN: + case GGML_OP_RWKV_WKV6: + case GGML_OP_RWKV_WKV7: { n_tasks = n_threads; } break; case GGML_OP_WIN_PART: case GGML_OP_WIN_UNPART: case GGML_OP_GET_REL_POS: - case GGML_OP_RWKV_WKV6: case GGML_OP_GATED_LINEAR_ATTN: case GGML_OP_MAP_UNARY: case GGML_OP_MAP_BINARY: diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 4f3578a3f..1f3bdb9b2 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -973,6 +973,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "GET_REL_POS", "ADD_REL_POS", "RWKV_WKV6", + "RWKV_WKV7", "GATED_LINEAR_ATTN", "UNARY", @@ -993,7 +994,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "OPT_STEP_ADAMW", }; -static_assert(GGML_OP_COUNT == 84, "GGML_OP_COUNT != 84"); +static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1071,6 +1072,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "get_rel_pos(x)", "add_rel_pos(x)", "rwkv_wkv6(k, v, r, tf, td, s)", + "rwkv_wkv7(r, w, k, v, a, b)", "gated_linear_attn(k, v, q, gate, s)", "unary(x)", @@ -1091,7 +1093,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "adamw(x)", }; -static_assert(GGML_OP_COUNT == 84, "GGML_OP_COUNT != 84"); +static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -4705,6 +4707,54 @@ struct ggml_tensor * ggml_rwkv_wkv6( return result; } +// ggml_rwkv_wkv7 + +struct ggml_tensor * ggml_rwkv_wkv7( + struct ggml_context * ctx, + struct ggml_tensor * r, + struct ggml_tensor * w, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * state) { + GGML_ASSERT(ggml_is_contiguous(r)); + GGML_ASSERT(ggml_is_contiguous(w)); + GGML_ASSERT(ggml_is_contiguous(k)); + GGML_ASSERT(ggml_is_contiguous(v)); + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(ggml_is_contiguous(b)); + GGML_ASSERT(ggml_is_contiguous(state)); + + const int64_t S = k->ne[0]; + const int64_t H = k->ne[1]; + const int64_t n_tokens = k->ne[2]; + const int64_t n_seqs = state->ne[1]; + { + GGML_ASSERT(w->ne[0] == S && w->ne[1] == H && w->ne[2] == n_tokens); + GGML_ASSERT(k->ne[0] == S && k->ne[1] == H && k->ne[2] == n_tokens); + GGML_ASSERT(v->ne[0] == S && v->ne[1] == H && v->ne[2] == n_tokens); + GGML_ASSERT(a->ne[0] == S && a->ne[1] == H && a->ne[2] == n_tokens); + GGML_ASSERT(b->ne[0] == S && b->ne[1] == H && b->ne[2] == n_tokens); + GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs); + } + + // concat output and new_state + const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + + result->op = GGML_OP_RWKV_WKV7; + result->src[0] = r; + result->src[1] = w; + result->src[2] = k; + result->src[3] = v; + result->src[4] = a; + result->src[5] = b; + result->src[6] = state; + + return result; +} + // ggml_gated_linear_attn struct ggml_tensor * ggml_gated_linear_attn( diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index ecac5b4bb..0adfe40b3 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -118,22 +118,26 @@ class Keys: TOKEN_SHIFT_COUNT = "{arch}.token_shift_count" class Attention: - HEAD_COUNT = "{arch}.attention.head_count" - HEAD_COUNT_KV = "{arch}.attention.head_count_kv" - MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" - CLAMP_KQV = "{arch}.attention.clamp_kqv" - KEY_LENGTH = "{arch}.attention.key_length" - VALUE_LENGTH = "{arch}.attention.value_length" - LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" - LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" - GROUPNORM_EPS = "{arch}.attention.group_norm_epsilon" - GROUPNORM_GROUPS = "{arch}.attention.group_norm_groups" - CAUSAL = "{arch}.attention.causal" - Q_LORA_RANK = "{arch}.attention.q_lora_rank" - KV_LORA_RANK = "{arch}.attention.kv_lora_rank" - REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" - SLIDING_WINDOW = "{arch}.attention.sliding_window" - SCALE = "{arch}.attention.scale" + HEAD_COUNT = "{arch}.attention.head_count" + HEAD_COUNT_KV = "{arch}.attention.head_count_kv" + MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" + CLAMP_KQV = "{arch}.attention.clamp_kqv" + KEY_LENGTH = "{arch}.attention.key_length" + VALUE_LENGTH = "{arch}.attention.value_length" + LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" + LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" + GROUPNORM_EPS = "{arch}.attention.group_norm_epsilon" + GROUPNORM_GROUPS = "{arch}.attention.group_norm_groups" + CAUSAL = "{arch}.attention.causal" + Q_LORA_RANK = "{arch}.attention.q_lora_rank" + KV_LORA_RANK = "{arch}.attention.kv_lora_rank" + DECAY_LORA_RANK = "{arch}.attention.decay_lora_rank" + ICLR_LORA_RANK = "{arch}.attention.iclr_lora_rank" + VALUE_RESIDUAL_MIX_LORA_RANK = "{arch}.attention.value_residual_mix_lora_rank" + GATE_LORA_RANK = "{arch}.attention.gate_lora_rank" + REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" + SLIDING_WINDOW = "{arch}.attention.sliding_window" + SCALE = "{arch}.attention.scale" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" @@ -256,6 +260,7 @@ class MODEL_ARCH(IntEnum): STARCODER2 = auto() RWKV6 = auto() RWKV6QWEN2 = auto() + RWKV7 = auto() MAMBA = auto() XVERSE = auto() COMMAND_R = auto() @@ -328,8 +333,20 @@ class MODEL_TENSOR(IntEnum): SSM_A = auto() SSM_D = auto() SSM_OUT = auto() + TIME_MIX_W0 = auto() TIME_MIX_W1 = auto() TIME_MIX_W2 = auto() + TIME_MIX_A0 = auto() + TIME_MIX_A1 = auto() + TIME_MIX_A2 = auto() + TIME_MIX_V0 = auto() + TIME_MIX_V1 = auto() + TIME_MIX_V2 = auto() + TIME_MIX_G1 = auto() + TIME_MIX_G2 = auto() + TIME_MIX_K_K = auto() + TIME_MIX_K_A = auto() + TIME_MIX_R_K = auto() TIME_MIX_LERP_X = auto() TIME_MIX_LERP_K = auto() TIME_MIX_LERP_V = auto() @@ -443,6 +460,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.STARCODER2: "starcoder2", MODEL_ARCH.RWKV6: "rwkv6", MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2", + MODEL_ARCH.RWKV7: "rwkv7", MODEL_ARCH.MAMBA: "mamba", MODEL_ARCH.XVERSE: "xverse", MODEL_ARCH.COMMAND_R: "command-r", @@ -515,8 +533,20 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a", MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0", MODEL_TENSOR.TIME_MIX_W1: "blk.{bid}.time_mix_w1", MODEL_TENSOR.TIME_MIX_W2: "blk.{bid}.time_mix_w2", + MODEL_TENSOR.TIME_MIX_A0: "blk.{bid}.time_mix_a0", + MODEL_TENSOR.TIME_MIX_A1: "blk.{bid}.time_mix_a1", + MODEL_TENSOR.TIME_MIX_A2: "blk.{bid}.time_mix_a2", + MODEL_TENSOR.TIME_MIX_V0: "blk.{bid}.time_mix_v0", + MODEL_TENSOR.TIME_MIX_V1: "blk.{bid}.time_mix_v1", + MODEL_TENSOR.TIME_MIX_V2: "blk.{bid}.time_mix_v2", + MODEL_TENSOR.TIME_MIX_G1: "blk.{bid}.time_mix_g1", + MODEL_TENSOR.TIME_MIX_G2: "blk.{bid}.time_mix_g2", + MODEL_TENSOR.TIME_MIX_K_K: "blk.{bid}.time_mix_k_k", + MODEL_TENSOR.TIME_MIX_K_A: "blk.{bid}.time_mix_k_a", + MODEL_TENSOR.TIME_MIX_R_K: "blk.{bid}.time_mix_r_k", MODEL_TENSOR.TIME_MIX_LERP_X: "blk.{bid}.time_mix_lerp_x", MODEL_TENSOR.TIME_MIX_LERP_K: "blk.{bid}.time_mix_lerp_k", MODEL_TENSOR.TIME_MIX_LERP_V: "blk.{bid}.time_mix_lerp_v", @@ -1153,6 +1183,37 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.RWKV7: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_W0, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_A0, + MODEL_TENSOR.TIME_MIX_A1, + MODEL_TENSOR.TIME_MIX_A2, + MODEL_TENSOR.TIME_MIX_V0, + MODEL_TENSOR.TIME_MIX_V1, + MODEL_TENSOR.TIME_MIX_V2, + MODEL_TENSOR.TIME_MIX_G1, + MODEL_TENSOR.TIME_MIX_G2, + MODEL_TENSOR.TIME_MIX_K_K, + MODEL_TENSOR.TIME_MIX_K_A, + MODEL_TENSOR.TIME_MIX_R_K, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.CHANNEL_MIX_LERP_K, + MODEL_TENSOR.CHANNEL_MIX_KEY, + MODEL_TENSOR.CHANNEL_MIX_VALUE, + ], MODEL_ARCH.MAMBA: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 080d2b9dc..af8b388df 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -767,6 +767,18 @@ class GGUFWriter: def add_kv_lora_rank(self, length: int) -> None: self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length) + def add_decay_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.DECAY_LORA_RANK.format(arch=self.arch), length) + + def add_iclr_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.ICLR_LORA_RANK.format(arch=self.arch), length) + + def add_value_residual_mix_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.VALUE_RESIDUAL_MIX_LORA_RANK.format(arch=self.arch), length) + + def add_gate_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.GATE_LORA_RANK.format(arch=self.arch), length) + def add_relative_attn_buckets_count(self, value: int) -> None: self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 617791e24..a3c56e780 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -27,7 +27,8 @@ class TensorNameMap: "embedding.word_embeddings", # chatglm "transformer.token_embeddings", # openelm "shared", # t5 - "rwkv.embeddings", # rwkv + "rwkv.embeddings", # rwkv v6 + "model.embeddings", # rwkv v7 ), # Token type embeddings @@ -41,7 +42,8 @@ class TensorNameMap: "embeddings.LayerNorm", # bert "emb_ln", # nomic-bert "transformer.norm", # openelm - "rwkv.blocks.0.pre_ln", # rwkv + "rwkv.blocks.0.pre_ln", # rwkv v6 + "model.pre_ln", # rwkv v7 "backbone.norm", # wavtokenizer ), @@ -81,7 +83,8 @@ class TensorNameMap: "encoder.final_layernorm", # chatglm "transformer.norm", # openelm "model.norm", # nemotron - "rwkv.ln_out", # rwkv + "rwkv.ln_out", # rwkv v6 + "model.ln_out", # rwkv v7 "backbone.final_layer_norm", # wavtokenizer ), @@ -122,14 +125,16 @@ class TensorNameMap: "transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx "encoder.layers.{bid}.input_layernorm", # chatglm "transformer.layers.{bid}.attn_norm", # openelm - "rwkv.blocks.{bid}.ln1", # rwkv + "rwkv.blocks.{bid}.ln1", # rwkv v6 + "model.blocks.{bid}.ln1", # rwkv v7 ), # Attention norm 2 MODEL_TENSOR.ATTN_NORM_2: ( "transformer.h.{bid}.ln_attn", # falcon40b "encoder.layer.{bid}.layer_norm_1", # jina-v2-code - "rwkv.blocks.{bid}.ln2", # rwkv + "rwkv.blocks.{bid}.ln2", # rwkv v6 + "model.blocks.{bid}.ln2", # rwkv v7 ), # Attention query-key-value @@ -462,14 +467,64 @@ class TensorNameMap: "backbone.layers.{bid}.mixer.out_proj", ), + MODEL_TENSOR.TIME_MIX_W0: ( + "model.blocks.{bid}.attention.w0", # rwkv7 + ), + MODEL_TENSOR.TIME_MIX_W1: ( "rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv v6 "model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2 + "model.blocks.{bid}.attention.w1" # rwkv7 ), MODEL_TENSOR.TIME_MIX_W2: ( "rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv v6 "model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2 + "model.blocks.{bid}.attention.w2" # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A0: ( + "model.blocks.{bid}.attention.a0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A1: ( + "model.blocks.{bid}.attention.a1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A2: ( + "model.blocks.{bid}.attention.a2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V0: ( + "model.blocks.{bid}.attention.v0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V1: ( + "model.blocks.{bid}.attention.v1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V2: ( + "model.blocks.{bid}.attention.v2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_G1: ( + "model.blocks.{bid}.attention.g1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_G2: ( + "model.blocks.{bid}.attention.g2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_K_K: ( + "model.blocks.{bid}.attention.k_k", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_K_A: ( + "model.blocks.{bid}.attention.k_a", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_R_K: ( + "model.blocks.{bid}.attention.r_k", # rwkv7 ), MODEL_TENSOR.TIME_MIX_LERP_X: ( @@ -522,36 +577,42 @@ class TensorNameMap: ), MODEL_TENSOR.TIME_MIX_KEY: ( - "rwkv.blocks.{bid}.attention.key", # rwkv + "rwkv.blocks.{bid}.attention.key", # rwkv v6 "model.layers.{bid}.self_attn.k_proj", # rwkv6qwen2 + "model.blocks.{bid}.attention.key", # rwkv v7 ), MODEL_TENSOR.TIME_MIX_VALUE: ( - "rwkv.blocks.{bid}.attention.value", # rwkv + "rwkv.blocks.{bid}.attention.value", # rwkv v6 "model.layers.{bid}.self_attn.v_proj", # rwkv6qwen2 + "model.blocks.{bid}.attention.value", # rwkv v7 ), MODEL_TENSOR.TIME_MIX_RECEPTANCE: ( - "rwkv.blocks.{bid}.attention.receptance", # rwkv - "model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2 + "rwkv.blocks.{bid}.attention.receptance", # rwkv v6 + "model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2 + "model.blocks.{bid}.attention.receptance", # rwkv v7 ), MODEL_TENSOR.TIME_MIX_GATE: ( - "rwkv.blocks.{bid}.attention.gate", # rwkv + "rwkv.blocks.{bid}.attention.gate", # rwkv v6 "model.layers.{bid}.self_attn.gate", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LN: ( - "rwkv.blocks.{bid}.attention.ln_x", # rwkv + "rwkv.blocks.{bid}.attention.ln_x", # rwkv v6 + "model.blocks.{bid}.attention.ln_x" # rwkv v7 ), MODEL_TENSOR.TIME_MIX_OUTPUT: ( "rwkv.blocks.{bid}.attention.output", # rwkv "model.layers.{bid}.self_attn.o_proj", # rwkv6qwen2 + "model.blocks.{bid}.attention.output", # rwkv v7 ), MODEL_TENSOR.CHANNEL_MIX_LERP_K: ( "rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv v6 + "model.blocks.{bid}.feed_forward.x_k", # rwkv v7 ), MODEL_TENSOR.CHANNEL_MIX_LERP_R: ( @@ -559,7 +620,8 @@ class TensorNameMap: ), MODEL_TENSOR.CHANNEL_MIX_KEY: ( - "rwkv.blocks.{bid}.feed_forward.key", # rwkv + "rwkv.blocks.{bid}.feed_forward.key", # rwkv v6 + "model.blocks.{bid}.feed_forward.key", # rwkv v7 ), MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: ( @@ -567,7 +629,8 @@ class TensorNameMap: ), MODEL_TENSOR.CHANNEL_MIX_VALUE: ( - "rwkv.blocks.{bid}.feed_forward.value", # rwkv + "rwkv.blocks.{bid}.feed_forward.value", # rwkv v6 + "model.blocks.{bid}.feed_forward.value", # rwkv v7 ), MODEL_TENSOR.ATTN_Q_A: ( diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 97a1e7e5e..72eeb7b52 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -58,6 +58,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_EXAONE, "exaone" }, { LLM_ARCH_RWKV6, "rwkv6" }, { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" }, + { LLM_ARCH_RWKV7, "rwkv7" }, { LLM_ARCH_GRANITE, "granite" }, { LLM_ARCH_GRANITE_MOE, "granitemoe" }, { LLM_ARCH_CHAMELEON, "chameleon" }, @@ -109,22 +110,26 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" }, { LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" }, - { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, - { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, - { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, - { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, - { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" }, - { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" }, - { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, - { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, - { LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" }, - { LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" }, - { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, - { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, - { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, - { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, - { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, - { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, + { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, + { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, + { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, + { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, + { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" }, + { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" }, + { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, + { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, + { LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" }, + { LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" }, + { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, + { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, + { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, + { LLM_KV_ATTENTION_DECAY_LORA_RANK, "%s.attention.decay_lora_rank" }, + { LLM_KV_ATTENTION_ICLR_LORA_RANK, "%s.attention.iclr_lora_rank" }, + { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" }, + { LLM_KV_ATTENTION_GATE_LORA_RANK, "%s.attention.gate_lora_rank" }, + { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, + { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, + { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, @@ -1217,6 +1222,40 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_RWKV7, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" }, + { LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" }, + { LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" }, + { LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" }, + { LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" }, + { LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" }, + { LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" }, + { LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" }, + { LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" }, + { LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" }, + { LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" }, + { LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" }, + { LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" }, + { LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" }, + { LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" }, + { LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" }, + { LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" }, + { LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" }, + { LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" }, + { LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" }, + { LLM_TENSOR_CHANNEL_MIX_LERP_K, "blk.%d.channel_mix_lerp_k" }, + { LLM_TENSOR_CHANNEL_MIX_KEY, "blk.%d.channel_mix_key" }, + { LLM_TENSOR_CHANNEL_MIX_VALUE, "blk.%d.channel_mix_value" }, + }, + }, { LLM_ARCH_GRANITE, { @@ -1376,6 +1415,12 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_A2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_V1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_V2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_G1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_G2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_DECAY_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_DECAY_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, @@ -1394,6 +1439,9 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_CHANNEL_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_TIME_MIX_K_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_TIME_MIX_K_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_TIME_MIX_R_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_TIME_MIX_LERP_W, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_LERP_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, @@ -1401,6 +1449,9 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_TIME_MIX_LERP_G, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_LERP_FUSED, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_DECAY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + {LLM_TENSOR_TIME_MIX_W0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + {LLM_TENSOR_TIME_MIX_A0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + {LLM_TENSOR_TIME_MIX_V0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}}, {LLM_TENSOR_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_ATTN_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, diff --git a/src/llama-arch.h b/src/llama-arch.h index 122fdcebe..193391e34 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -62,6 +62,7 @@ enum llm_arch { LLM_ARCH_EXAONE, LLM_ARCH_RWKV6, LLM_ARCH_RWKV6QWEN2, + LLM_ARCH_RWKV7, LLM_ARCH_GRANITE, LLM_ARCH_GRANITE_MOE, LLM_ARCH_CHAMELEON, @@ -126,6 +127,10 @@ enum llm_kv { LLM_KV_ATTENTION_CAUSAL, LLM_KV_ATTENTION_Q_LORA_RANK, LLM_KV_ATTENTION_KV_LORA_RANK, + LLM_KV_ATTENTION_DECAY_LORA_RANK, + LLM_KV_ATTENTION_ICLR_LORA_RANK, + LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, + LLM_KV_ATTENTION_GATE_LORA_RANK, LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, LLM_KV_ATTENTION_SLIDING_WINDOW, LLM_KV_ATTENTION_SCALE, @@ -249,8 +254,20 @@ enum llm_tensor { LLM_TENSOR_SSM_A, LLM_TENSOR_SSM_D, LLM_TENSOR_SSM_OUT, + LLM_TENSOR_TIME_MIX_W0, LLM_TENSOR_TIME_MIX_W1, LLM_TENSOR_TIME_MIX_W2, + LLM_TENSOR_TIME_MIX_A0, + LLM_TENSOR_TIME_MIX_A1, + LLM_TENSOR_TIME_MIX_A2, + LLM_TENSOR_TIME_MIX_V0, + LLM_TENSOR_TIME_MIX_V1, + LLM_TENSOR_TIME_MIX_V2, + LLM_TENSOR_TIME_MIX_G1, + LLM_TENSOR_TIME_MIX_G2, + LLM_TENSOR_TIME_MIX_K_K, + LLM_TENSOR_TIME_MIX_K_A, + LLM_TENSOR_TIME_MIX_R_K, LLM_TENSOR_TIME_MIX_LERP_X, LLM_TENSOR_TIME_MIX_LERP_W, LLM_TENSOR_TIME_MIX_LERP_K, diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 1fe454103..1b3044e11 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -75,6 +75,10 @@ struct llama_hparams { uint32_t time_decay_extra_dim = 0; uint32_t wkv_head_size = 0; uint32_t token_shift_count = 2; + uint32_t n_lora_decay = 0; + uint32_t n_lora_iclr = 0; + uint32_t n_lora_value_res_mix = 0; + uint32_t n_lora_gate = 0; float rope_attn_factor = 1.0f; float rope_freq_base_train; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 0f4b62c43..50fd51e12 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1210,6 +1210,21 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_RWKV7: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_WKV_HEAD_SIZE, hparams.wkv_head_size); + ml.get_key(LLM_KV_ATTENTION_DECAY_LORA_RANK, hparams.n_lora_decay); + ml.get_key(LLM_KV_ATTENTION_ICLR_LORA_RANK, hparams.n_lora_iclr); + ml.get_key(LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, hparams.n_lora_value_res_mix); + ml.get_key(LLM_KV_ATTENTION_GATE_LORA_RANK, hparams.n_lora_gate); + ml.get_key(LLM_KV_TOKEN_SHIFT_COUNT, hparams.token_shift_count, false); + + switch (hparams.n_layer) { + // TODO: Add variants + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: { @@ -3280,6 +3295,78 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; + case LLM_ARCH_RWKV7: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // Block 0, LN0 + tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0); + tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + const int n_lora_decay = hparams.n_lora_decay; + const int n_lora_iclr = hparams.n_lora_iclr; + const int n_lora_value_res_mix = hparams.n_lora_value_res_mix; + const int n_lora_gate = hparams.n_lora_gate; + const int attn_hidden_size = n_embd; + const int ffn_size = hparams.n_ff_arr[0]; + + 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.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0); + + layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, 0); + layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, 0); + + layer.time_mix_w0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W0, "weight", i), {n_embd}, 0); + layer.time_mix_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W1, "weight", i), {n_embd, n_lora_decay}, 0); + layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {n_lora_decay, n_embd}, 0); + + layer.time_mix_a0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A0, "weight", i), {n_embd}, 0); + layer.time_mix_a1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A1, "weight", i), {n_embd, n_lora_iclr}, 0); + layer.time_mix_a2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A2, "weight", i), {n_lora_iclr, n_embd}, 0); + + if (i == 0) { + // actually not used + layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0); + layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_iclr}, 0); + layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_iclr, n_embd}, 0); + } else { + layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0); + layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_value_res_mix}, 0); + layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0); + } + + layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, 0); + layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, 0); + + layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 6}, 0); + + layer.time_mix_k_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_K, "weight", i), {attn_hidden_size}, 0); + layer.time_mix_k_a = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_A, "weight", i), {attn_hidden_size}, 0); + layer.time_mix_r_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_R_K, "weight", i), {attn_hidden_size}, 0); + + layer.time_mix_key = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "weight", i), {attn_hidden_size, n_embd}, 0); + layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0); + layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0); + + layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, 0); + layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, 0); + layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0); + + layer.channel_mix_lerp_k = create_tensor(tn(LLM_TENSOR_CHANNEL_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, 0); + + layer.channel_mix_key = create_tensor(tn(LLM_TENSOR_CHANNEL_MIX_KEY, "weight", i), {n_embd, ffn_size}, 0); + layer.channel_mix_value = create_tensor(tn(LLM_TENSOR_CHANNEL_MIX_VALUE, "weight", i), {ffn_size, n_embd}, 0); + } + + } break; case LLM_ARCH_CHAMELEON: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -3865,6 +3952,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { case LLM_ARCH_JAIS: case LLM_ARCH_RWKV6: case LLM_ARCH_RWKV6QWEN2: + case LLM_ARCH_RWKV7: case LLM_ARCH_WAVTOKENIZER_DEC: return LLAMA_ROPE_TYPE_NONE; @@ -4018,6 +4106,7 @@ bool llama_model_is_recurrent(const struct llama_model * model) { case LLM_ARCH_MAMBA: return true; case LLM_ARCH_RWKV6: return true; case LLM_ARCH_RWKV6QWEN2: return true; + case LLM_ARCH_RWKV7: return true; default: return false; } } diff --git a/src/llama-model.h b/src/llama-model.h index a7c304447..697b97e9b 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -256,6 +256,20 @@ struct llama_layer { struct ggml_tensor * time_mix_receptance_b = nullptr; struct ggml_tensor * time_mix_gate = nullptr; + // rwkv v7 + struct ggml_tensor * time_mix_w0 = nullptr; + struct ggml_tensor * time_mix_a0 = nullptr; + struct ggml_tensor * time_mix_a1 = nullptr; + struct ggml_tensor * time_mix_a2 = nullptr; + struct ggml_tensor * time_mix_v0 = nullptr; + struct ggml_tensor * time_mix_v1 = nullptr; + struct ggml_tensor * time_mix_v2 = nullptr; + struct ggml_tensor * time_mix_g1 = nullptr; + struct ggml_tensor * time_mix_g2 = nullptr; + struct ggml_tensor * time_mix_k_k = nullptr; + struct ggml_tensor * time_mix_k_a = nullptr; + struct ggml_tensor * time_mix_r_k = nullptr; + struct ggml_tensor * time_mix_ln = nullptr; struct ggml_tensor * time_mix_ln_b = nullptr; struct ggml_tensor * time_mix_output = nullptr; diff --git a/src/llama.cpp b/src/llama.cpp index 607f27861..de802ce3d 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -1039,6 +1039,127 @@ static struct ggml_tensor * llm_build_rwkv6_channel_mix( return ggml_mul(ctx, r, llm_build_lora_mm(lctx, ctx, layer->channel_mix_value, k)); } +static struct ggml_tensor * llm_build_rwkv7_time_mix( + struct llama_context & lctx, + struct ggml_context * ctx, + const struct llama_layer * layer, + struct ggml_tensor * cur, + struct ggml_tensor * x_prev, + struct ggml_tensor ** wkv_state, + struct ggml_tensor * & first_layer_value, + size_t wkv_head_size) { + size_t n_embd = cur->ne[0]; + size_t n_seq_tokens = cur->ne[1]; + size_t n_seqs = cur->ne[2]; + + size_t head_size = wkv_head_size; + size_t head_count = n_embd / head_size; + + size_t n_tokens = n_seqs * n_seq_tokens; + + struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur); + struct ggml_tensor * dummy = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, n_tokens, 6); + sx = ggml_repeat(ctx, sx, dummy); + + struct ggml_tensor * xxx = ggml_add(ctx, ggml_mul(ctx, sx, layer->time_mix_lerp_fused), cur); + + struct ggml_tensor * xr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0); + struct ggml_tensor * xw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + struct ggml_tensor * xk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + struct ggml_tensor * xv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + struct ggml_tensor * xa = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + struct ggml_tensor * xg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 5 * sizeof(float)); + + struct ggml_tensor * r = llm_build_lora_mm(lctx, ctx, layer->time_mix_receptance, xr); + // Assume that there won't be lora adapters on these “lora” matmuls? + struct ggml_tensor * w = ggml_add( + ctx, + ggml_mul_mat(ctx, layer->time_mix_w2, ggml_tanh(ctx, ggml_mul_mat(ctx, layer->time_mix_w1, xw))), + layer->time_mix_w0 + ); + w = ggml_exp(ctx, ggml_scale(ctx, ggml_sigmoid(ctx, w), -0.606531)); + + struct ggml_tensor * k = llm_build_lora_mm(lctx, ctx, layer->time_mix_key, xk); + struct ggml_tensor * v = llm_build_lora_mm(lctx, ctx, layer->time_mix_value, xv); + if (first_layer_value == nullptr) { + first_layer_value = v; + } else { + // Add the first layer value as a residual connection. + v = ggml_add(ctx, v, + ggml_mul(ctx, + ggml_sub(ctx, first_layer_value, v), + ggml_sigmoid(ctx, ggml_add(ctx, + ggml_mul_mat(ctx, layer->time_mix_v2, ggml_mul_mat(ctx, layer->time_mix_v1, xv)), + layer->time_mix_v0 + ) + ) + ) + ); + } + + struct ggml_tensor * g = ggml_mul_mat(ctx, layer->time_mix_g2, ggml_sigmoid(ctx, ggml_mul_mat(ctx, layer->time_mix_g1, xg))); + struct ggml_tensor * a = ggml_sigmoid(ctx, + ggml_add( + ctx, + ggml_mul_mat(ctx, layer->time_mix_a2, ggml_mul_mat(ctx, layer->time_mix_a1, xa)), + layer->time_mix_a0 + ) + ); + + struct ggml_tensor * kk = ggml_reshape_3d(ctx, ggml_mul(ctx, k, layer->time_mix_k_k), head_size, head_count, n_tokens); + kk = ggml_l2_norm(ctx, kk, 1e-12); + + struct ggml_tensor * ka = ggml_mul(ctx, k, layer->time_mix_k_a); + k = ggml_add(ctx, k, ggml_sub(ctx, ggml_mul(ctx, a, ka), ka)); + + r = ggml_reshape_3d(ctx, r, head_size, head_count, n_tokens); + w = ggml_reshape_3d(ctx, w, head_size, head_count, n_tokens); + k = ggml_reshape_3d(ctx, k, head_size, head_count, n_tokens); + v = ggml_reshape_3d(ctx, v, head_size, head_count, n_tokens); + a = ggml_reshape_3d(ctx, a, head_size, head_count, n_tokens); + + struct ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx, r, w, k, v, ggml_neg(ctx, kk), ggml_mul(ctx, kk, a), *wkv_state); + cur = ggml_view_1d(ctx, wkv_output, n_embd * n_tokens, 0); + *wkv_state = ggml_view_1d(ctx, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); + + // group norm with head_count groups + cur = ggml_reshape_3d(ctx, cur, n_embd / head_count, head_count, n_tokens); + cur = ggml_norm(ctx, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens); + cur = ggml_add(ctx, ggml_mul(ctx, cur, layer->time_mix_ln), layer->time_mix_ln_b); + + struct ggml_tensor * rk = ggml_sum_rows(ctx, + ggml_mul(ctx, ggml_mul(ctx, k, r), ggml_reshape_2d(ctx, layer->time_mix_r_k, head_size, head_count))); + cur = ggml_add(ctx, cur, ggml_reshape_2d(ctx, ggml_mul(ctx, v, rk), n_embd, n_tokens)); + + cur = ggml_mul(ctx, cur, g); + cur = llm_build_lora_mm(lctx, ctx, layer->time_mix_output, cur); + + return cur; +} + +static struct ggml_tensor * llm_build_rwkv7_channel_mix( + struct llama_context & lctx, + struct ggml_context * ctx, + const struct llama_layer * layer, + struct ggml_tensor * cur, + struct ggml_tensor * x_prev) { + struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur); + struct ggml_tensor * xk = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_k), cur); + + struct ggml_tensor * k = ggml_sqr( + ctx, + ggml_relu( + ctx, + llm_build_lora_mm(lctx, ctx, layer->channel_mix_key, xk) + ) + ); + + return llm_build_lora_mm(lctx, ctx, layer->channel_mix_value, k); +} + struct llm_build_context { const llama_model & model; llama_context & lctx; @@ -7781,6 +7902,119 @@ struct llm_build_context { return gf; } + ggml_cgraph * build_rwkv7() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); + + GGML_ASSERT(n_embd == hparams.n_embd_k_s() / hparams.token_shift_count); + + const int64_t n_seqs = ubatch.n_seqs; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + const int64_t n_tokens = ubatch.n_tokens; + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs); + GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + struct ggml_tensor * state_copy = build_inp_s_copy(); + struct ggml_tensor * state_mask = build_inp_s_mask(); + struct ggml_tensor * value_first_layer = nullptr; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); + inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + + // (ab)using the KV cache to store the states + struct ggml_tensor * token_shift = llm_build_copy_mask_state(ctx0, + gf, kv_self.k_l[il], state_copy, state_mask, + hparams.n_embd_k_s(), kv_self.size, kv_head, n_kv, n_seqs); + struct ggml_tensor * wkv_states = llm_build_copy_mask_state(ctx0, + gf, kv_self.v_l[il], state_copy, state_mask, + hparams.n_embd_v_s(), kv_self.size, kv_head, n_kv, n_seqs); + + cur = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + token_shift = ggml_reshape_3d(ctx0, token_shift, n_embd, 2, n_seqs); + + struct ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + struct ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + struct ggml_tensor * x_norm_att = llm_build_norm(ctx0, cur, hparams, layer->attn_norm, layer->attn_norm_b, LLM_NORM, cb, il); + struct ggml_tensor * x_prev = ggml_concat( + ctx0, + att_shift, + ggml_view_3d(ctx0, x_norm_att, n_embd, n_seq_tokens - 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], 0), + 1 + ); + + cur = ggml_add(ctx0, cur, llm_build_rwkv7_time_mix(lctx, ctx0, layer, x_norm_att, x_prev, &wkv_states, value_first_layer, hparams.wkv_head_size)); + ggml_build_forward_expand(gf, cur); + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + wkv_states, + ggml_view_1d( + ctx0, + kv_self.v_l[il], + hparams.n_embd_v_s() * n_seqs, + hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il]) + ) + ) + ); + + struct ggml_tensor * x_norm_ffn = llm_build_norm(ctx0, cur, hparams, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, cb, il); + x_prev = ggml_concat( + ctx0, + ffn_shift, + ggml_view_3d(ctx0, x_norm_ffn, n_embd, n_seq_tokens - 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], 0), + 1 + ); + cur = ggml_add(ctx0, cur, llm_build_rwkv7_channel_mix(lctx, ctx0, layer, x_norm_ffn, x_prev)); + ggml_build_forward_expand(gf, cur); + + struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att)); + struct ggml_tensor * last_norm_ffn = ggml_view_3d(ctx0, x_norm_ffn, n_embd, 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_ffn)); + + token_shift = ggml_concat(ctx0, last_norm_att, last_norm_ffn, 1); + + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * 2, 0), + ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il])) + ) + ); + + if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { + cur = ggml_scale(ctx0, cur, 0.5F); + } + + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + + cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + // ref: https://github.com/facebookresearch/chameleon // based on the original build_llama() function, changes: // * qk-norm @@ -8394,6 +8628,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_rwkv6qwen2(); } break; + case LLM_ARCH_RWKV7: + { + result = llm.build_rwkv7(); + } break; case LLM_ARCH_CHAMELEON: { result = llm.build_chameleon(); diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 13846caf6..5418dea48 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1868,6 +1868,37 @@ struct test_rwkv_wkv6 : public test_case { } }; +// GGML_OP_RWKV_WKV7 +struct test_rwkv_wkv7 : public test_case { + const ggml_type type; + + const int64_t head_count; + const int64_t head_size; + const int64_t n_seq_tokens; + const int64_t n_seqs; + + std::string vars() override { + return VARS_TO_STR5(type, head_count, head_size, n_seq_tokens, n_seqs); + } + + test_rwkv_wkv7(ggml_type type = GGML_TYPE_F32, + int64_t head_count = 32, int64_t head_size = 64, int64_t n_seq_tokens = 32, int64_t n_seqs = 32) + : type(type), head_count(head_count), head_size(head_size), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + const int64_t n_tokens = n_seq_tokens * n_seqs; + ggml_tensor * r = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data()); + ggml_tensor * w = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data()); + ggml_tensor * k = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data()); + ggml_tensor * v = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data()); + ggml_tensor * a = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data()); + ggml_tensor * b = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data()); + ggml_tensor * s = ggml_new_tensor(ctx, type, 2, std::vector{ head_size * head_size * head_count, n_seqs }.data()); + ggml_tensor * out = ggml_rwkv_wkv7(ctx, r, w, k, v, a, b, s); + return out; + } +}; + // GGML_OP_GATED_LINEAR_ATTN struct test_gla : public test_case { const ggml_type type; @@ -4026,6 +4057,11 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 4)); test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 128, 4)); + test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 1, 1)); + test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 1)); + test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 4)); + test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 128, 4)); + test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 1, 1)); test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 1)); test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));