issue/263 merge T2-1-4

Co-authored-by: PanZezhong <panzezhong@qiyuanlab.com>

include/infinicore_infer.h

@@ -4,7 +4,11 @@
 #include "infinicore_infer/cache.h"
 #include "infinicore_infer/weights_loader.h"
 
+
 #include "infinicore_infer/models/deepseek.h"
 #include "infinicore_infer/models/jiuge.h"
+#include "infinicore_infer/models/jiuge_awq.h"
+#include "infinicore_infer/models/qwen3vl.h"
+
 
 #endif /* INFINICORE_INFER_H */

include/infinicore_infer/models/qwen3vl.h

+#ifndef QWEN3VL_WEIGHTS_H
+#define QWEN3VL_WEIGHTS_H
+
+#include <infiniccl.h>
+#include <infiniop.h>
+#include <infinirt.h>
+
+#include <stddef.h>
+#include <stdint.h>
+
+struct Qwen3vlWeights;
+
+// Function pointer signatures
+typedef void (*qwen3vl_load_global_fn)(Qwen3vlWeights *, void *cpu_ptr);
+typedef void (*qwen3vl_load_layer_fn)(Qwen3vlWeights *, void *cpu_ptr, size_t layer_id);
+// Struct containing all weight loading functions
+typedef struct {
+    // Global
+    qwen3vl_load_global_fn load_input_embd;
+    qwen3vl_load_global_fn load_output_norm;
+    qwen3vl_load_global_fn load_output_embd;
+
+    // Attention
+    qwen3vl_load_layer_fn load_attn_norm;
+    qwen3vl_load_layer_fn load_attn_q_norm;
+    qwen3vl_load_layer_fn load_attn_k_norm;
+    qwen3vl_load_layer_fn load_attn_qkv_proj;
+    qwen3vl_load_layer_fn load_attn_o_proj;
+    
+    // MLP
+    qwen3vl_load_layer_fn load_mlp_norm;
+    qwen3vl_load_layer_fn load_mlp_gate_up;
+    qwen3vl_load_layer_fn load_mlp_down;
+
+} Qwen3vlLangWeightLoader;
+
+typedef struct {
+    // Patch_embed
+    qwen3vl_load_global_fn load_patch_embed_weight;
+    qwen3vl_load_global_fn load_patch_embed_bias;
+    qwen3vl_load_global_fn load_pos_embed_weight;
+
+    // blocks attn
+    qwen3vl_load_layer_fn load_attn_proj_weight;
+    qwen3vl_load_layer_fn load_attn_proj_bias;
+    qwen3vl_load_layer_fn load_attn_qkv_weight;
+    qwen3vl_load_layer_fn load_attn_qkv_bias;
+
+    //block mlp
+    qwen3vl_load_layer_fn load_mlp_linear_fc1_weight;
+    qwen3vl_load_layer_fn load_mlp_linear_fc1_bias;
+    qwen3vl_load_layer_fn load_mlp_linear_fc2_weight;
+    qwen3vl_load_layer_fn load_mlp_linear_fc2_bias;
+
+    //block norm
+    qwen3vl_load_layer_fn  load_norm1_weight;
+    qwen3vl_load_layer_fn  load_norm1_bias;
+    qwen3vl_load_layer_fn  load_norm2_weight;
+    qwen3vl_load_layer_fn  load_norm2_bias;
+
+    //deepstack_merger
+    qwen3vl_load_layer_fn load_deepstack_merger_linear_fc1_weight;
+    qwen3vl_load_layer_fn load_deepstack_merger_linear_fc1_bias;
+    qwen3vl_load_layer_fn load_deepstack_merger_linear_fc2_weight;
+    qwen3vl_load_layer_fn load_deepstack_merger_linear_fc2_bias;
+    qwen3vl_load_layer_fn load_deepstack_merger_norm_weight;
+    qwen3vl_load_layer_fn load_deepstack_merger_norm_bias;
+
+    //merger
+    qwen3vl_load_global_fn load_merger_linear_fc1_weight;
+    qwen3vl_load_global_fn load_merger_linear_fc1_bias;
+    qwen3vl_load_global_fn load_merger_linear_fc2_weight;
+    qwen3vl_load_global_fn load_merger_linear_fc2_bias;
+    qwen3vl_load_global_fn load_merger_norm_weight;
+    qwen3vl_load_global_fn load_merger_norm_bias;
+
+} Qwen3vlVisWeightLoader;
+
+typedef struct { 
+    Qwen3vlLangWeightLoader lang_loader;
+    Qwen3vlVisWeightLoader vis_loader;
+} Qwen3vlWeightLoader;
+
+struct Qwen3vlModel;
+
+typedef struct {
+    size_t bos_token_id;
+    size_t eos_token_id;
+    size_t head_dim;
+    size_t hidden_size;
+    float initializer_range;
+    size_t intermediate_size;
+    size_t max_tokens;
+    size_t num_attention_heads;
+    size_t num_hidden_layers;
+    size_t num_key_value_heads;
+    float rms_norm_eps;
+    size_t mrope_section[3];
+    size_t rope_theta;
+    size_t vocab_size;
+} Qwen3vlTextMeta;
+
+typedef struct {
+    size_t depth;
+    size_t deepstack_visual_indexes[3];
+    size_t hidden_size;
+    size_t in_channels;
+    float initializer_range;
+    size_t intermediate_size;
+    size_t num_heads;
+    size_t num_position_embeddings;
+    size_t out_hidden_size;
+    size_t patch_size;
+    size_t spatial_merge_size;
+    size_t temporal_patch_size;
+} Qwen3vlVisMeta;
+
+typedef struct {
+    infiniDtype_t dtype; //INFINI_DTYPE_BF16
+
+    Qwen3vlTextMeta text_meta;
+    Qwen3vlVisMeta vis_meta;
+
+    size_t image_token_id;
+    size_t video_token_id;
+    size_t vision_end_token_id;
+    size_t vision_start_token_id;
+} Qwen3vlMeta;
+
+//////////////////// APIs ///////////////////////
+/// @brief 创建模型
+/// @param device 协处理器种类
+/// @param ndev 协处理器数量
+/// @param dev_ids 协处理器编号，长度为 ndev
+__C __export struct Qwen3vlModel *
+createQwen3vlModel(const Qwen3vlMeta *,
+                      const Qwen3vlWeights *);
+
+__C Qwen3vlWeights *
+createQwen3vlWeights(const Qwen3vlMeta *meta,
+                        infiniDevice_t device,
+                        int ndev,
+                        const int *dev_ids,
+                        bool transpose_weight);
+
+__C __export Qwen3vlWeightLoader *
+createQwen3vlWeightLoader();
+
+/// @brief 销毁模型
+__C __export void destroyQwen3vlModel(struct Qwen3vlModel *);
+
+__C __export struct Qwen3vlCache *
+createQwen3vlCache(const struct Qwen3vlModel *);
+
+__C __export void
+dropQwen3vlCache(const struct Qwen3vlModel *,
+                    struct Qwen3vlCache *);
+
+/// @brief 批次推理一轮，并采样出新的 token
+/// @param tokens 输入 token 地址
+/// @param ntok 输入 token 数量
+/// @param nreq 请求数量
+/// @param req_lens 每个请求的 token 数量
+/// @param req_pos 每个请求的起始位置
+/// @param kv_caches 每个请求的 KV Cache
+/// @param temperature 采样温度（0. 表示贪心采样）
+/// @param topk 采样 topk（1 表示贪心采样）
+/// @param topp 采样 topp
+/// @param output 输出 token 数组，每个请求一个输出，长度至少为nreq
+__C __export void
+inferBatchQwen3vl(struct Qwen3vlModel *,
+                    const uint32_t *tokens, uint32_t ntok,
+                    void *pixel_values, uint32_t total_patches,
+                    uint32_t *image_grid_thw, uint32_t num_images,
+                    void *pixel_values_videos, uint32_t total_patches_videos,
+                    uint32_t *video_grid_thw, uint32_t num_videos,
+                    uint32_t patch_features,
+                    const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos,
+                    struct Qwen3vlCache **caches,
+                    const float *temperature, const uint32_t *topk, const float *topp,
+                    uint32_t *output);
+
+/// @brief 批次推理一轮，输出 output embedding 后的 logits
+/// @param tokens 输入 token 地址
+/// @param ntok 输入 token 数量
+/// @param nreq 请求数量
+/// @param req_lens 每个请求的 token 数量
+/// @param req_pos 每个请求的起始位置
+/// @param kv_caches 每个请求的 KV Cache
+/// @param logits 输出 token 数组，每个请求一个输出，长度至少为nreq
+__C __export void
+forwardBatchQwen3vl(struct Qwen3vlModel *,
+                    const uint32_t *tokens, uint32_t ntok,
+                    void *pixel_values, uint32_t total_patches,
+                    uint32_t *image_grid_thw, uint32_t num_images,
+                    void *pixel_values_videos, uint32_t total_patches_videos,
+                    uint32_t *video_grid_thw, uint32_t num_videos,
+                    uint32_t patch_features,
+                    const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos,
+                    struct Qwen3vlCache **caches,
+                    void *logits);
+
+#endif // QWEN3VL_WEIGHTS_H

scripts/libinfinicore_infer/__init__.py

@@ -6,7 +6,17 @@ from .deepseek_v3 import (
     DeepSeekV3MetaCStruct,
     DeepSeekV3WeightsCStruct,
     DeepSeekV3WeightLoaderCStruct,
-    DeepSeekV3CacheCStruct,
+)
+from .qwen3vl import (
+    Qwen3vlModel,
+    Qwen3vlMetaCStruct,
+    TextMetaCStruct,
+    VisMetaCStruct,
+    Qwen3vlWeightsCStruct,
+    Qwen3vlWeightLoaderCStruct,
+    Qwen3vlVisWeightLoaderCStruct,
+    Qwen3vlLangWeightLoaderCStruct,
+    Qwen3vlCacheCStruct,
 )
 
 __all__ = [
@@ -23,5 +33,14 @@ __all__ = [
     "DeepSeekV3MetaCStruct",
     "DeepSeekV3WeightsCStruct",
     "DeepSeekV3WeightLoaderCStruct",
+    "Qwen3vlModel",
+    "Qwen3vlMetaCStruct",
+    "TextMetaCStruct",
+    "VisMetaCStruct",
+    "Qwen3vlWeightsCStruct",
+    "Qwen3vlWeightLoaderCStruct",
+    "Qwen3vlVisWeightLoaderCStruct",
+    "Qwen3vlLangWeightLoaderCStruct",
+    "Qwen3vlCacheCStruct",
     "ModelRegister",
 ]

scripts/libinfinicore_infer/qwen3vl.py

+from .base import BaseModel, DataType, DeviceType, KVCacheCStruct, register_model
+from ctypes import (
+    c_size_t,
+    c_uint,
+    c_uint16,
+    c_int,
+    c_float,
+    c_void_p,
+    c_bool,
+    POINTER,
+    Structure,
+    CFUNCTYPE,
+)
+
+
+class TextMetaCStruct(Structure):
+    _fields_ = [
+        ("bos_token_id", c_size_t),
+        ("eos_token_id", c_size_t),
+        ("head_dim", c_size_t),
+        ("hidden_size", c_size_t),
+        ("initializer_range", c_float),
+        ("_pad1", c_float),
+        ("intermediate_size", c_size_t),
+        ("max_tokens", c_size_t),
+        ("num_attention_heads", c_size_t),
+        ("num_hidden_layers", c_size_t),
+        ("num_key_value_heads", c_size_t),
+        ("rms_norm_eps", c_float),
+        ("_pad2", c_float),
+        ("mrope_section", c_size_t * 3),
+        ("rope_theta", c_size_t),
+        ("vocab_size", c_size_t),
+    ]
+
+
+class VisMetaCStruct(Structure):
+    _fields_ = [
+        ("depth", c_size_t),
+        ("deepstack_visual_indexes", c_size_t * 3),
+        ("hidden_size", c_size_t),
+        ("in_channels", c_size_t),
+        ("initializer_range", c_float),
+        ("_pad1", c_float),
+        ("intermediate_size", c_size_t),
+        ("num_heads", c_size_t),
+        ("num_position_embeddings", c_size_t),
+        ("out_hidden_size", c_size_t),
+        ("patch_size", c_size_t),
+        ("spatial_merge_size", c_size_t),
+        ("temporal_patch_size", c_size_t),
+    ]
+
+
+class Qwen3vlMetaCStruct(Structure):
+    _fields_ = [
+        ("dtype", DataType),
+        ("_pad_dtype", c_uint),
+        ("text_meta", TextMetaCStruct),
+        ("vis_meta", VisMetaCStruct),
+        # Token ids
+        ("image_token_id", c_size_t),
+        ("video_token_id", c_size_t),
+        ("vision_end_token_id", c_size_t),
+        ("vision_start_token_id", c_size_t),
+    ]
+
+
+class Qwen3vlWeightsCStruct(Structure):
+    pass
+
+
+class Qwen3vlModelCStruct(Structure):
+    pass
+
+
+class Qwen3vlCacheCStruct(Structure):
+    pass
+
+
+load_global_fn = CFUNCTYPE(None, POINTER(Qwen3vlWeightsCStruct), c_void_p)
+load_layer_fn = CFUNCTYPE(None, POINTER(Qwen3vlWeightsCStruct), c_void_p, c_size_t)
+
+
+class Qwen3vlLangWeightLoaderCStruct(Structure):
+    _fields_ = [
+        # Global
+        ("load_input_embd", load_global_fn),
+        ("load_output_norm", load_global_fn),
+        ("load_output_embd", load_global_fn),
+        # Attention
+        ("load_attn_norm", load_layer_fn),
+        ("load_attn_q_norm", load_layer_fn),
+        ("load_attn_k_norm", load_layer_fn),
+        ("load_attn_qkv_proj", load_layer_fn),
+        ("load_attn_o_proj", load_layer_fn),
+        # MLP
+        ("load_mlp_norm", load_layer_fn),
+        ("load_mlp_gate_up", load_layer_fn),
+        ("load_mlp_down", load_layer_fn),
+    ]
+
+
+class Qwen3vlVisWeightLoaderCStruct(Structure):
+    _fields_ = [
+        # Patch embed
+        ("load_patch_embed_weight", load_global_fn),
+        ("load_patch_embed_bias", load_global_fn),
+        ("load_pos_embed_weight", load_global_fn),
+        # Blocks attention
+        ("load_attn_proj_weight", load_layer_fn),
+        ("load_attn_proj_bias", load_layer_fn),
+        ("load_attn_qkv_weight", load_layer_fn),
+        ("load_attn_qkv_bias", load_layer_fn),
+        # Blocks MLP
+        ("load_mlp_linear_fc1_weight", load_layer_fn),
+        ("load_mlp_linear_fc1_bias", load_layer_fn),
+        ("load_mlp_linear_fc2_weight", load_layer_fn),
+        ("load_mlp_linear_fc2_bias", load_layer_fn),
+        # Blocks norm
+        ("load_norm1_weight", load_layer_fn),
+        ("load_norm1_bias", load_layer_fn),
+        ("load_norm2_weight", load_layer_fn),
+        ("load_norm2_bias", load_layer_fn),
+        # Deepstack merger
+        ("load_deepstack_merger_linear_fc1_weight", load_layer_fn),
+        ("load_deepstack_merger_linear_fc1_bias", load_layer_fn),
+        ("load_deepstack_merger_linear_fc2_weight", load_layer_fn),
+        ("load_deepstack_merger_linear_fc2_bias", load_layer_fn),
+        ("load_deepstack_merger_norm_weight", load_layer_fn),
+        ("load_deepstack_merger_norm_bias", load_layer_fn),
+        # Merger
+        ("load_merger_linear_fc1_weight", load_global_fn),
+        ("load_merger_linear_fc1_bias", load_global_fn),
+        ("load_merger_linear_fc2_weight", load_global_fn),
+        ("load_merger_linear_fc2_bias", load_global_fn),
+        ("load_merger_norm_weight", load_global_fn),
+        ("load_merger_norm_bias", load_global_fn),
+    ]
+
+
+class Qwen3vlWeightLoaderCStruct(Structure):
+    _fields_ = [
+        ("lang_loader", Qwen3vlLangWeightLoaderCStruct),
+        ("vis_loader", Qwen3vlVisWeightLoaderCStruct),
+    ]
+
+
+@register_model
+class Qwen3vlModel(BaseModel):
+    @classmethod
+    def register_lib(cls, lib):
+        """Register Qwen3vl model functions with the library"""
+        lib.createQwen3vlWeightLoader.argtypes = []
+        lib.createQwen3vlWeightLoader.restype = POINTER(Qwen3vlWeightLoaderCStruct)
+
+        lib.createQwen3vlWeights.argtypes = [
+            POINTER(Qwen3vlMetaCStruct),
+            DeviceType,
+            c_int,
+            POINTER(c_int),
+            c_bool,
+        ]
+        lib.createQwen3vlWeights.restype = POINTER(Qwen3vlWeightsCStruct)
+
+        lib.createQwen3vlModel.argtypes = [
+            POINTER(Qwen3vlMetaCStruct),
+            POINTER(Qwen3vlWeightsCStruct),
+        ]
+        lib.createQwen3vlModel.restype = POINTER(Qwen3vlModelCStruct)
+
+        lib.destroyQwen3vlModel.argtypes = [POINTER(Qwen3vlModelCStruct)]
+
+        lib.createQwen3vlCache.argtypes = [POINTER(Qwen3vlModelCStruct)]
+        lib.createQwen3vlCache.restype = POINTER(Qwen3vlCacheCStruct)
+
+        lib.dropQwen3vlCache.argtypes = [
+            POINTER(Qwen3vlModelCStruct),
+            POINTER(Qwen3vlCacheCStruct),
+        ]
+
+        lib.inferBatchQwen3vl.argtypes = [
+            POINTER(Qwen3vlModelCStruct),
+            POINTER(c_uint),
+            c_uint,
+            c_void_p,          # pixel_values,
+            c_uint,            # total_patches,
+            POINTER(c_uint),   # image_grid_thw,
+            c_uint,            # num_images,
+            c_void_p,          # pixel_values_videos,
+            c_uint,            # total_patches_videos,
+            POINTER(c_uint),   # video_grid_thw,
+            c_uint,            # num_videos,
+            c_uint,            # patch_features,
+            POINTER(c_uint),
+            c_uint,
+            POINTER(c_uint),
+            POINTER(POINTER(Qwen3vlCacheCStruct)),
+            POINTER(c_float),
+            POINTER(c_uint),
+            POINTER(c_float),
+            POINTER(c_uint),
+        ]
+
+        lib.forwardBatchQwen3vl.argtypes = [
+            POINTER(Qwen3vlModelCStruct),
+            POINTER(c_uint),
+            c_uint,
+            c_void_p,          # pixel_values,
+            c_uint,            # total_patches,
+            POINTER(c_uint),   # image_grid_thw,
+            c_uint,            # num_images,
+            c_void_p,          # pixel_values_videos,
+            c_uint,            # total_patches_videos,
+            POINTER(c_uint),   # video_grid_thw,
+            c_uint,            # num_videos,
+            c_uint,            # patch_features,
+            POINTER(c_uint),
+            c_uint,
+            POINTER(c_uint),
+            POINTER(POINTER(Qwen3vlCacheCStruct)),
+            c_void_p,
+        ]
+
+    def create_weight_loader(self):
+        return self.lib.createQwen3vlWeightLoader()
+
+    def create_weights(self, meta, device_type, ndev, dev_ids, transpose_weight):
+        return self.lib.createQwen3vlWeights(meta, device_type, ndev, dev_ids, transpose_weight)
+
+    def create_model(self, meta, weights):
+        return self.lib.createQwen3vlModel(meta, weights)
+
+    def destroy_model(self, model):
+        self.lib.destroyQwen3vlModel(model)
+
+    def create_cache(self, model):
+        return self.lib.createQwen3vlCache(model)
+
+    def drop_cache(self, model, cache):
+        self.lib.dropQwen3vlCache(model, cache)
+
+    def infer_batch(
+        self,
+        model,
+        tokens,
+        ntok,
+        pixel_values,
+        total_patches,
+        image_grid_thw,
+        num_images,
+        pixel_values_videos,
+        total_patches_videos,
+        video_grid_thw,
+        num_videos,
+        patch_features,
+        req_lens,
+        nreq,
+        req_pos,
+        caches,
+        temperature,
+        topk,
+        topp,
+        output,
+    ):
+        self.lib.inferBatchQwen3vl(
+            model,
+            tokens,
+            ntok,
+            pixel_values,
+            total_patches,
+            image_grid_thw,
+            num_images,
+            pixel_values_videos,
+            total_patches_videos,
+            video_grid_thw,
+            num_videos,
+            patch_features,
+            req_lens,
+            nreq,
+            req_pos,
+            caches,
+            temperature,
+            topk,
+            topp,
+            output,
+        )
+
+    def forward_batch(
+        self,
+        model,
+        tokens,
+        ntok,
+        pixel_values,
+        total_patches,
+        image_grid_thw,
+        num_images,
+        pixel_values_videos,
+        total_patches_videos,
+        video_grid_thw,
+        num_videos,
+        patch_features,
+        req_lens,
+        nreq,
+        req_pos,
+        caches,
+        logits,
+    ):
+        self.lib.forwardBatchQwen3vl(
+            model,
+            tokens,
+            ntok,
+            pixel_values,
+            total_patches,
+            image_grid_thw,
+            num_images,
+            pixel_values_videos,
+            total_patches_videos,
+            video_grid_thw,
+            num_videos,
+            patch_features,
+            req_lens,
+            nreq,
+            req_pos,
+            caches,
+            logits,
+        )
\ No newline at end of file

src/cache_manager/opcache_manager.hpp

@@ -153,6 +153,8 @@ public:
 class CacheManager {
 public:
     DECLARE_OP_CACHE(Add)
+    DECLARE_OP_CACHE(Conv)
+    DECLARE_OP_CACHE(Mul)
     DECLARE_OP_CACHE(RMSNorm)
     DECLARE_OP_CACHE(Gemm)
     DECLARE_OP_CACHE(RoPE)
@@ -160,11 +162,14 @@ public:
     DECLARE_OP_CACHE(CausalSoftmax)
     DECLARE_OP_CACHE(Topkrouter)
     DECLARE_OP_CACHE(SwiGLU)
+    DECLARE_OP_CACHE(Silu)
     DECLARE_OP_CACHE(RandomSample)
     DECLARE_OP_CACHE(DequantizeAWQ)
 
     CacheManager(size_t capacity = 100)
         : Add_cache(capacity, DESTROY_FUNC(Add)),
+          Conv_cache(capacity, DESTROY_FUNC(Conv)),
+          Mul_cache(capacity, DESTROY_FUNC(Mul)),
           RMSNorm_cache(capacity, DESTROY_FUNC(RMSNorm)),
           Gemm_cache(capacity, DESTROY_FUNC(Gemm)),
           RoPE_cache(capacity, DESTROY_FUNC(RoPE)),
@@ -172,6 +177,7 @@ public:
           CausalSoftmax_cache(capacity, DESTROY_FUNC(CausalSoftmax)),
           Topkrouter_cache(capacity, DESTROY_FUNC(Topkrouter)),
           SwiGLU_cache(capacity, DESTROY_FUNC(SwiGLU)),
+          Silu_cache(capacity, DESTROY_FUNC(Silu)),
           RandomSample_cache(capacity, DESTROY_FUNC(RandomSample)),
           DequantizeAWQ_cache(capacity, DESTROY_FUNC(DequantizeAWQ)) {}
 

src/models/inference_context.cpp

@@ -33,6 +33,61 @@ void InferenceContext::add(std::shared_ptr<Tensor> c,
         c->data(), a->data(), b->data(), stream));
 }
 
+void InferenceContext::conv(std::shared_ptr<Tensor> y,
+                             std::shared_ptr<Tensor> x,
+                             std::shared_ptr<Tensor> w,
+                             std::shared_ptr<Tensor> bias,
+                             void *pads,
+                             void *strides,
+                             void *dilations,
+                             size_t n) {
+    size_t key = CacheManager::createDescriptorKey(y, x, w, bias);
+    // Combine additional parameters into the key for unique identification
+    hash_combine(key, std::hash<void*>()(pads));
+    hash_combine(key, std::hash<void*>()(strides));
+    hash_combine(key, std::hash<void*>()(dilations));
+    hash_combine(key, std::hash<size_t>()(n));
+
+    infiniopConvDescriptor_t desc;
+    if (!cache_manager->getConvDescriptor(key, desc)) {
+        RUN_INFINI(infiniopCreateConvDescriptor(
+            op_handle, &desc, y->desc(), x->desc(), w->desc(), 
+            bias ? bias->desc() : nullptr, pads, strides, dilations, n));
+        cache_manager->putConvDescriptor(key, desc);
+    }
+
+    size_t workspace_size = 0;
+    RUN_INFINI(infiniopGetConvWorkspaceSize(desc, &workspace_size));
+    ensure_workspace(workspace_size);
+    void *workspace = workspace_storage->memory();
+
+    RUN_INFINI(infiniopConv(
+        desc, workspace, workspace_size,
+        y->data(), x->data(), w->data(), 
+        bias ? bias->data() : nullptr, stream));
+}
+
+void InferenceContext::mul(std::shared_ptr<Tensor> c,
+                           std::shared_ptr<Tensor> a,
+                           std::shared_ptr<Tensor> b) {
+    size_t key = CacheManager::createDescriptorKey(c, a, b);
+
+    infiniopMulDescriptor_t desc;
+    if (!cache_manager->getMulDescriptor(key, desc)) {
+        RUN_INFINI(infiniopCreateMulDescriptor(op_handle, &desc, c->desc(), a->desc(), b->desc()));
+        cache_manager->putMulDescriptor(key, desc);
+    }
+
+    size_t workspace_size = 0;
+    RUN_INFINI(infiniopGetMulWorkspaceSize(desc, &workspace_size));
+    ensure_workspace(workspace_size);
+    void *workspace = workspace_storage->memory();
+
+    RUN_INFINI(infiniopMul(
+        desc, workspace, workspace_size,
+        c->data(), a->data(), b->data(), stream));
+}
+
 void InferenceContext::rmsnorm(std::shared_ptr<Tensor> y,
                                std::shared_ptr<Tensor> x,
                                std::shared_ptr<Tensor> w,
@@ -189,6 +244,26 @@ void InferenceContext::swiglu(std::shared_ptr<Tensor> out,
                               out->data(), up->data(), gate->data(), stream));
 }
 
+void InferenceContext::silu(std::shared_ptr<Tensor> out,
+                            std::shared_ptr<Tensor> input) {
+    size_t key = CacheManager::createDescriptorKey(out, input);
+
+    infiniopSiluDescriptor_t desc;
+    if (!cache_manager->getSiluDescriptor(key, desc)) {
+        RUN_INFINI(infiniopCreateSiluDescriptor(
+            op_handle, &desc, out->desc(), input->desc()));
+        cache_manager->putSiluDescriptor(key, desc);
+    }
+
+    size_t workspace_size = 0;
+    RUN_INFINI(infiniopGetSiluWorkspaceSize(desc, &workspace_size));
+    ensure_workspace(workspace_size);
+    void *workspace = workspace_storage->memory();
+
+    RUN_INFINI(infiniopSilu(desc, workspace, workspace_size,
+                            out->data(), input->data(), stream));
+}
+
 void InferenceContext::randomSample(std::shared_ptr<Tensor> out,
                                     std::shared_ptr<Tensor> prob,
                                     float random_val, float top_p, uint32_t top_k, float temperature) {

src/models/inference_context.hpp

@@ -19,6 +19,14 @@ struct InferenceContext {
     void add(std::shared_ptr<Tensor> c,
              std::shared_ptr<Tensor> a,
              std::shared_ptr<Tensor> b);
+    void conv(std::shared_ptr<Tensor> y,
+              std::shared_ptr<Tensor> x,
+              std::shared_ptr<Tensor> w,
+              std::shared_ptr<Tensor> bias,
+              void *pads, void *strides, void *dilations, size_t n);
+    void mul(std::shared_ptr<Tensor> c,
+             std::shared_ptr<Tensor> a,
+             std::shared_ptr<Tensor> b);
     void rmsnorm(std::shared_ptr<Tensor> y,
                  std::shared_ptr<Tensor> x,
                  std::shared_ptr<Tensor> w,
@@ -48,6 +56,8 @@ struct InferenceContext {
     void swiglu(std::shared_ptr<Tensor> out,
                 std::shared_ptr<Tensor> up,
                 std::shared_ptr<Tensor> gate);
+    void silu(std::shared_ptr<Tensor> out,
+              std::shared_ptr<Tensor> input);
     void randomSample(std::shared_ptr<Tensor> out,
                       std::shared_ptr<Tensor> prob,
                       float random_val, float top_p, uint32_t top_k, float temperature);
@@ -81,6 +91,15 @@ inline void add(std::shared_ptr<Tensor> c, std::shared_ptr<Tensor> a, std::share
     getInferenceContext().add(c, a, b);
 }
 
+inline void conv(std::shared_ptr<Tensor> y, std::shared_ptr<Tensor> x, std::shared_ptr<Tensor> w, std::shared_ptr<Tensor> bias,
+                void *pads, void *strides, void *dilations, size_t n) {
+    getInferenceContext().conv(y, x, w, bias, pads, strides, dilations, n);
+}
+
+inline void mul(std::shared_ptr<Tensor> c, std::shared_ptr<Tensor> a, std::shared_ptr<Tensor> b) {
+    getInferenceContext().mul(c, a, b);
+}
+
 inline void rmsnorm(std::shared_ptr<Tensor> y, std::shared_ptr<Tensor> x,
                     std::shared_ptr<Tensor> w, float epsilon) {
     getInferenceContext().rmsnorm(y, x, w, epsilon);
@@ -131,6 +150,10 @@ inline void swiglu(std::shared_ptr<Tensor> out, std::shared_ptr<Tensor> up,
     getInferenceContext().swiglu(out, up, gate);
 }
 
+inline void silu(std::shared_ptr<Tensor> out, std::shared_ptr<Tensor> input) {
+    getInferenceContext().silu(out, input);
+}
+
 inline void randomSample(std::shared_ptr<Tensor> out, std::shared_ptr<Tensor> prob,
                          float random_val, float top_p, uint32_t top_k, float temperature) {
     getInferenceContext().randomSample(out, prob, random_val, top_p, top_k, temperature);

src/models/qwen3vl/qwen3vl_cache.cpp

+#include "qwen3vl_impl.hpp"
+
+__C struct Qwen3vlCache *
+createQwen3vlCache(const struct Qwen3vlModel *model) {
+    Qwen3vlCache *cache = new Qwen3vlCache();
+    auto ndev = model->dev_resources.size();
+    auto nlayer = model->meta.text_meta.num_hidden_layers;
+    auto max_len = model->meta.text_meta.max_tokens;
+    auto dh = model->meta.text_meta.head_dim;
+    auto nkv = model->meta.text_meta.num_key_value_heads / size_t(ndev);
+    auto k_rot_shape = std::vector<size_t>{max_len, nkv, dh};
+    auto v_shape = std::vector<size_t>{max_len, nkv, dh};
+    for (size_t idev = 0; idev < ndev; idev++) {
+        RUN_INFINI(infinirtSetDevice(model->device, model->dev_ids[idev]));
+        auto k_rot_cache = std::vector<std::shared_ptr<Tensor>>();
+        auto v_cache = std::vector<std::shared_ptr<Tensor>>();
+        for (size_t layer = 0; layer < nlayer; layer++) {
+            k_rot_cache.push_back(std::move(Tensor::buffer(model->meta.dtype, k_rot_shape)));
+            v_cache.push_back(std::move(Tensor::buffer(model->meta.dtype, v_shape)));
+        }
+        cache->k_rot.push_back(k_rot_cache);
+        cache->v.push_back(v_cache);
+    }
+
+    return cache;
+}
+
+//////还有visual deepstack需要cache？
+
+__C void
+dropQwen3vlCache(const struct Qwen3vlModel *model,
+                    struct Qwen3vlCache *cache) {
+    auto ndev = model->dev_resources.size();
+    auto nlayer = model->meta.text_meta.num_hidden_layers;
+    for (size_t idev = 0; idev < ndev; idev++) {
+        RUN_INFINI(infinirtSetDevice(model->device, model->dev_ids[idev]));
+        for (size_t layer = 0; layer < nlayer; layer++) {
+            cache->k_rot[idev][layer].reset();
+            cache->v[idev][layer].reset();
+        }
+    }
+    delete cache;
+}
\ No newline at end of file

src/models/qwen3vl/qwen3vl_impl.hpp

+#ifndef QWEN3VL_IMPL_H
+#define QWEN3VL_IMPL_H
+
+#include "infinicore_infer.h"
+
+#include "../../allocator.hpp"
+#include "../../tensor.hpp"
+
+#include <condition_variable>
+#include <memory>
+#include <mutex>
+#include <thread>
+#include <vector>
+
+struct Qwen3vlLayerWeight {
+    std::shared_ptr<Tensor> attn_norm;
+    std::shared_ptr<Tensor> attn_qkv_proj;
+    std::shared_ptr<Tensor> attn_q_norm;
+    std::shared_ptr<Tensor> attn_k_norm;
+    std::shared_ptr<Tensor> attn_o_proj;
+
+    std::shared_ptr<Tensor> mlp_norm;
+    std::shared_ptr<Tensor> mlp_gate_up;
+    std::shared_ptr<Tensor> mlp_down;
+};
+
+struct Qwen3vlLanguageModelWeight {
+    std::shared_ptr<Tensor> in_embd, out_embd, out_norm;
+    std::vector<Qwen3vlLayerWeight> layers;
+};
+
+struct Qwen3vlVisBlockWeight {
+    std::shared_ptr<Tensor> attn_proj_weight, attn_proj_bias, attn_qkv_weight, attn_qkv_bias;
+    std::shared_ptr<Tensor> mlp_linear_fc1_weight, mlp_linear_fc1_bias, mlp_linear_fc2_weight, mlp_linear_fc2_bias;
+    std::shared_ptr<Tensor> norm1_weight, norm1_bias, norm2_weight, norm2_bias;
+};
+
+struct DeepstackMergerWeight {
+    std::shared_ptr<Tensor> linear_fc1_weight, linear_fc1_bias, linear_fc2_weight, linear_fc2_bias;
+    std::shared_ptr<Tensor> norm_weight, norm_bias;
+};
+
+struct MergerWeight {
+    std::shared_ptr<Tensor> linear_fc1_weight, linear_fc1_bias, linear_fc2_weight, linear_fc2_bias;
+    std::shared_ptr<Tensor> norm_weight, norm_bias;
+};
+
+
+struct Qwen3vlVisualEncoderWeight {
+    std::shared_ptr<Tensor> patch_embed_weight, patch_embed_bias, pos_embed_weight;
+    std::vector<Qwen3vlVisBlockWeight> blocks;
+    std::vector<DeepstackMergerWeight> deepstack_mergers;
+    std::shared_ptr<MergerWeight> merger;
+};
+
+
+struct Qwen3vlDeviceWeights {
+    std::shared_ptr<Tensor> sin_table,cos_table;
+    std::shared_ptr<Qwen3vlLanguageModelWeight> w_lang;
+    std::shared_ptr<Qwen3vlVisualEncoderWeight> w_vis;
+    infiniDevice_t device;
+    int dev_id;
+    infinirtStream_t load_stream;
+};
+
+struct Qwen3vlWeights {
+    Qwen3vlMeta const *meta;
+    bool transpose_weight;
+    std::vector<std::shared_ptr<Qwen3vlDeviceWeights>> device_weights;
+
+    Qwen3vlWeights(const Qwen3vlMeta *meta,
+                      infiniDevice_t device,
+                      int ndev,
+                      const int *dev_ids,
+                      bool transpose_weight);
+};
+
+struct Qwen3vlDeviceResource {
+    // Device
+    infiniDevice_t device;
+    int device_id;
+    infiniopHandle_t handle;
+    // Weights
+    std::shared_ptr<Qwen3vlDeviceWeights> weights;
+    // Streams
+    infinirtStream_t stream;
+    // Communicator
+    infinicclComm_t comm;
+
+    std::shared_ptr<MemoryPool> memory_pool;
+};
+
+struct InferState { // qwen3vl namespace
+    inline static std::mutex mtx_sync;
+    inline static int sync_cnt;
+    inline static std::condition_variable cv_sync;
+    std::mutex mtx;
+    std::condition_variable cv_load, cv_start, cv_done;
+    bool loaded = false;
+    bool proceed = false;
+    bool exit_flag = false;
+};
+
+struct InferRequest { // qwen3vl namespace
+    const uint32_t *tokens;
+    uint32_t ntok;
+    void *pixel_values;
+    uint32_t total_patches;
+    uint32_t *image_grid_thw;
+    uint32_t num_images;
+    void *pixel_values_videos;
+    uint32_t total_patches_videos;
+    uint32_t *video_grid_thw;
+    uint32_t num_videos;
+    uint32_t patch_features;
+    const uint32_t *req_lens;
+    uint32_t nreq;
+    const uint32_t *req_pos;
+    struct Qwen3vlCache **kv_caches;
+    const float *temperature;
+    const uint32_t *topk;
+    const float *topp;
+    uint32_t *output;
+    void *logits;
+};
+
+struct Qwen3vlModel {
+    Qwen3vlMeta meta;
+    infiniDevice_t device;
+    std::vector<int> dev_ids;
+    std::vector<Qwen3vlDeviceResource> dev_resources;
+    std::vector<InferState> states;
+    std::vector<std::thread> threads;
+    InferRequest req;
+
+    Qwen3vlModel(const Qwen3vlMeta *, const Qwen3vlWeights *weights);
+};
+
+struct Qwen3vlCache {
+    std::vector<std::vector<std::shared_ptr<Tensor>>> k_rot, v; 
+};
+
+#endif
\ No newline at end of file

src/tensor/tensor.cpp

@@ -267,7 +267,7 @@ void print_data_bf16(uint16_t const *data, const std::vector<size_t> &shape,
         std::cout << std::endl;
     } else if (dim < shape.size() - 1) {
         for (size_t i = 0; i < shape[dim]; i++) {
-            print_data(data + i * strides[dim], shape, strides, dim + 1);
+            print_data_bf16(data + i * strides[dim], shape, strides, dim + 1);
         }
     }
 }