llama: add support for QRWKV6 model architecture (#11001)

llama: add support for QRWKV6 model architecture (#11001) * WIP: Add support for RWKV6Qwen2 Signed-off-by: Molly Sophia <[email protected]> * RWKV: Some graph simplification Signed-off-by: Molly Sophia <[email protected]> * Add support for RWKV6Qwen2 with cpu and cuda GLA Signed-off-by: Molly Sophia <[email protected]> * RWKV6[QWEN2]: Concat lerp weights together to reduce cpu overhead Signed-off-by: Molly Sophia <[email protected]> * Fix some typos Signed-off-by: Molly Sophia <[email protected]> * code format changes Signed-off-by: Molly Sophia <[email protected]> * Fix wkv test & add gla test Signed-off-by: Molly Sophia <[email protected]> * Fix cuda warning Signed-off-by: Molly Sophia <[email protected]> * Update README.md Signed-off-by: Molly Sophia <[email protected]> * Update ggml/src/ggml-cuda/gla.cu Co-authored-by: Georgi Gerganov <[email protected]> * Fix fused lerp weights loading with RWKV6 Signed-off-by: Molly Sophia <[email protected]> * better sanity check skipping for QRWKV6 in llama-quant thanks @compilade Signed-off-by: Molly Sophia <[email protected]> Co-authored-by: compilade <[email protected]> --------- Signed-off-by: Molly Sophia <[email protected]> Co-authored-by: Georgi Gerganov <[email protected]> Co-authored-by: compilade <[email protected]>
ggerganov · Jan 10, 2025 · ee7136c · ee7136c
1 parent c6860cc
commit ee7136c
Show file tree

Hide file tree

Showing 23 changed files with 863 additions and 125 deletions.
diff --git a/README.md b/README.md
@@ -99,6 +99,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
+- [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 
 #### Multimodal

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
@@ -326,6 +326,7 @@ def prepare_tensors(self):
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
                             gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
+                            gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
                             gguf.MODEL_TENSOR.POSNET_NORM1,
                             gguf.MODEL_TENSOR.POSNET_NORM2,
                         )
@@ -3316,6 +3317,8 @@ def set_gguf_parameters(self):
         # 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]]:
         new_name = self.map_tensor_name(name)
 
@@ -3331,14 +3334,84 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         if new_name.endswith("time_mix_decay.weight") or "lerp" in new_name:
             data_torch = data_torch.squeeze()
 
-        rescale_every_n_layers = self.hparams["rescale_every"]
-        if rescale_every_n_layers > 0:
-            if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
-                data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
+        try:
+            rescale_every_n_layers = self.hparams["rescale_every"]
+            if rescale_every_n_layers > 0:
+                if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
+                    data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
+        except KeyError:
+            pass
+
+        # concat time_mix_lerp weights to reduce some cpu overhead
+        # also reduces the number of tensors in the model
+        if bid is not None and "time_mix_lerp" in new_name and "time_mix_lerp_x" not in new_name:
+            try:
+                self.lerp_weights[bid][new_name] = data_torch
+            except KeyError:
+                self.lerp_weights[bid] = {new_name: data_torch}
+            if all(f"blk.{bid}.time_mix_lerp_{i}.weight" in self.lerp_weights[bid].keys() for i in ["w", "k", "v", "r", "g"]):
+                new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                data = torch.stack([self.lerp_weights[bid][f"blk.{bid}.time_mix_lerp_{i}.weight"].unsqueeze(0) for i in ["w", "k", "v", "r", "g"]], dim=0).unsqueeze(1)
+                yield (new_name, data)
+            return
 
         yield (new_name, data_torch)
 
 
+@Model.register("RWKV6Qwen2ForCausalLM")
+class RWKV6Qwen2Model(Rwkv6Model):
+    model_arch = gguf.MODEL_ARCH.RWKV6QWEN2
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        num_attention_heads = self.hparams["num_attention_heads"]
+        num_key_value_heads = self.hparams["num_key_value_heads"]
+        hidden_size = self.hparams["hidden_size"]
+        head_size = hidden_size // num_attention_heads
+        rms_norm_eps = self.hparams["rms_norm_eps"]
+        intermediate_size = self.hparams["intermediate_size"]
+        time_mix_extra_dim = 64 if hidden_size >= 4096 else 32
+        time_decay_extra_dim = 128 if hidden_size >= 4096 else 64
+
+        # 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_wkv_head_size(head_size)
+        self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
+        self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # special parameters for time_mixing in RWKV6QWEN2
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_token_shift_count(1)
+        # RWKV6QWEN2 use grouped key/value like GQA
+        self.gguf_writer.add_head_count_kv(num_key_value_heads)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        for new_name, data in super().modify_tensors(data_torch, name, bid):
+            if "time_mix_w1" in new_name or "time_mix_w2" in new_name:
+                data = data.view(5, -1, data.shape[-1])
+                # rwkv6qwen2 has a different order of rkvwg instead of the original wkvrg
+                # permute them here to avoid code changes
+                data = torch.stack([data[3], data[1], data[2], data[0], data[4]], dim=0).view(-1, data.shape[-1])
+                if "w2" in new_name:
+                    data = data.view(5, -1, data.shape[-1])
+                yield (new_name, data)
+                continue
+            yield (new_name, data)
+
+
 @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
@@ -501,6 +501,7 @@ extern "C" {
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
         GGML_OP_RWKV_WKV6,
+        GGML_OP_GATED_LINEAR_ATTN,
 
         GGML_OP_UNARY,
 
@@ -1859,6 +1860,15 @@ extern "C" {
             struct ggml_tensor  * td,
             struct ggml_tensor  * state);
 
+    GGML_API struct ggml_tensor * ggml_gated_linear_attn(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * g,
+            struct ggml_tensor  * state,
+            float scale);
+
     // custom operators
 
     typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);

diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -11803,9 +11803,9 @@ static void ggml_compute_forward_add_rel_pos(
 static void ggml_compute_forward_rwkv_wkv6_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
-    const int64_t T = dst->src[1]->ne[3];
+    const int64_t T = dst->src[1]->ne[2];
     const int64_t C = dst->ne[0];
-    const int64_t HEADS = dst->src[1]->ne[2];
+    const int64_t HEADS = dst->src[1]->ne[1];
     const int64_t n_seqs = dst->src[5]->ne[1];
     const int64_t head_size = C / HEADS;
 
@@ -12000,6 +12000,197 @@ static void ggml_compute_forward_rwkv_wkv6(
     }
 }
 
+// ggml_compute_forward_gla
+
+static void ggml_compute_forward_gla_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[4]->ne[1];
+    const int64_t head_size = C / HEADS;
+    const float scale = ggml_get_op_params_f32(dst, 0);
+
+    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 * k = (float *) dst->src[0]->data;
+    float * v = (float *) dst->src[1]->data;
+    float * q = (float *) dst->src[2]->data;
+    float * g = (float *) dst->src[3]->data;
+
+    size_t t_stride = HEADS * head_size; // Same to C
+
+    size_t h_stride = C / HEADS;
+    GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS
+    size_t h_stride_2d = head_size * head_size;
+
+    if (ith == 0) {
+        memset(dst_data, 0, T * C * sizeof(float));
+    }
+    ggml_barrier(params->threadpool);
+
+
+    #if defined(__AVX__) && !defined(__AVX512F__)
+        #define GGML_F32X GGML_F32x8
+        #define GGML_F32X_SET1 GGML_F32x8_SET1
+        #define GGML_F32X_LOAD GGML_F32x8_LOAD
+        #define GGML_F32X_STORE GGML_F32x8_STORE
+        #define GGML_F32X_MUL GGML_F32x8_MUL
+        #define GGML_F32X_FMA GGML_F32x8_FMA
+        #define GLA_VECTOR_SIZE 8
+    #elif defined(__AVX512F__)
+        #define GGML_F32X GGML_F32x16
+        #define GGML_F32X_SET1 GGML_F32x16_SET1
+        #define GGML_F32X_LOAD GGML_F32x16_LOAD
+        #define GGML_F32X_STORE GGML_F32x16_STORE
+        #define GGML_F32X_MUL GGML_F32x16_MUL
+        #define GGML_F32X_FMA GGML_F32x16_FMA
+        #define GLA_VECTOR_SIZE 16
+    #elif defined(__ARM_NEON) && defined(__aarch64__)
+        #define GGML_F32X GGML_F32x4
+        #define GGML_F32X_SET1 GGML_F32x4_SET1
+        #define GGML_F32X_LOAD GGML_F32x4_LOAD
+        #define GGML_F32X_STORE GGML_F32x4_STORE
+        #define GGML_F32X_MUL GGML_F32x4_MUL
+        #define GGML_F32X_FMA GGML_F32x4_FMA
+        #define GLA_VECTOR_SIZE 4
+    #endif
+
+    #ifdef GLA_VECTOR_SIZE
+        const int64_t vec_count = head_size / GLA_VECTOR_SIZE;
+
+        for (int64_t t = 0; t < T; t++) {
+            size_t t_offset = t * t_stride;
+            size_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[4]->data + state_offset;
+
+            for (int64_t h = h_start; h < h_end; h++) {
+                size_t h_offset = h * h_stride;
+                size_t t_h_offset = t_offset + h_offset;
+                size_t h_2d_offset = h * h_stride_2d;
+
+                for (int64_t i = 0; i < head_size; i++) {
+                    size_t t_h_i_offset = t_h_offset + i;
+                    size_t h_2d_i_offset = h_2d_offset + i * h_stride;
+
+                    float k_val = k[t_h_i_offset];
+                    float q_val = q[t_h_i_offset] * scale;
+                    float g_val = g[t_h_i_offset];
+
+                    // Broadcast scalar values to vectors
+                    GGML_F32X k_vec = GGML_F32X_SET1(k_val);
+                    GGML_F32X q_vec = GGML_F32X_SET1(q_val);
+                    GGML_F32X g_vec = GGML_F32X_SET1(g_val);
+
+                    for (int64_t j = 0; j < vec_count; j++) {
+                        size_t base_j = j * GLA_VECTOR_SIZE;
+                        size_t t_h_j_offset = t_h_offset + base_j;
+                        size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
+
+                        // Load x elements at once
+                        GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]);
+                        GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]);
+                        GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]);
+
+                        // Compute kv = v * k
+                        GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec);
+
+                        // Compute temp = prev_state * g + kv
+                        GGML_F32X temp_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, g_vec);
+
+                        // Update dst: dst += temp * q
+                        dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, q_vec);
+                        GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec);
+
+                        // Update state
+                        GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], temp_vec);
+                    }
+
+                    // Handle remaining elements, this will not be used.
+                    for (int64_t j = vec_count * GLA_VECTOR_SIZE; j < head_size; j++) {
+                        size_t t_h_j_offset = t_h_offset + j;
+                        size_t h_2d_i_j_offset = h_2d_i_offset + j;
+                        float v_val = v[t_h_j_offset];
+                        float kv_val = v_val * k_val;
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        float temp_val = kv_val + prev_state_val * g_val;
+                        dst_data[t_h_j_offset] += temp_val * q_val;
+                        state_cur[h_2d_i_j_offset] = temp_val;
+                    }
+                }
+            }
+        }
+
+    #else
+        for (int64_t t = 0; t < T; t++) {
+            size_t t_offset = t * t_stride;
+            size_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[4]->data + state_offset;
+
+            for (int64_t h = h_start; h < h_end; h++) {
+                size_t h_offset = h * h_stride;
+                size_t t_h_offset = t_offset + h_offset;
+                size_t h_2d_offset = h * h_stride_2d;
+
+                for (int64_t i = 0; i < head_size; i++) {
+                    size_t t_h_i_offset = t_h_offset + i;
+                    size_t h_2d_i_offset = h_2d_offset + i * h_stride;
+
+                    float k_val = k[t_h_i_offset];
+                    float q_val = q[t_h_i_offset] * scale;
+                    float g_val = g[t_h_i_offset];
+
+                    for (int64_t j = 0; j < head_size; j++) {
+                        size_t t_h_j_offset = t_h_offset + j;
+                        size_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                        float v_val = v[t_h_j_offset];
+                        float kv_val = v_val * k_val;
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        float temp_val = prev_state_val * g_val + kv_val;
+                        dst_data[t_h_j_offset] += temp_val * q_val;
+                        state_cur[h_2d_i_j_offset] = temp_val;
+                    }
+                }
+            }
+        }
+    #endif
+}
+
+
+static void ggml_compute_forward_gla(
+        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_gla_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_map_unary
 
 static void ggml_compute_forward_map_unary_f32(
@@ -12749,6 +12940,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv6(params, tensor);
             } break;
+        case GGML_OP_GATED_LINEAR_ATTN:
+            {
+                ggml_compute_forward_gla(params, tensor);
+            } break;
         case GGML_OP_MAP_UNARY:
             {
                 ggml_unary_op_f32_t fun;
@@ -13047,6 +13242,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         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:
         case GGML_OP_MAP_CUSTOM1_F32:

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -37,6 +37,7 @@
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv6.cuh"
+#include "ggml-cuda/gla.cuh"
 
 #include <algorithm>
 #include <array>
@@ -2167,6 +2168,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_RWKV_WKV6:
             ggml_cuda_op_rwkv_wkv6(ctx, dst);
             break;
+        case GGML_OP_GATED_LINEAR_ATTN:
+            ggml_cuda_op_gated_linear_attn(ctx, dst);
+            break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             ggml_cuda_cross_entropy_loss_back(ctx, dst);
             break;
@@ -3011,6 +3015,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_RWKV_WKV6:
+        case GGML_OP_GATED_LINEAR_ATTN:
             return true;
         case GGML_OP_FLASH_ATTN_EXT: {
 #ifndef FLASH_ATTN_AVAILABLE