Merge pull request #1 from afeldman-nm/enc_dec_t5

T5 enc/dec example file; linting/formatting

Author: js8544

README.md

@@ -78,6 +78,7 @@ vLLM seamlessly supports many Hugging Face models, including the following archi
 - Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
 - Qwen2 (`Qwen/Qwen2-7B-beta`, `Qwen/Qwen-7B-Chat-beta`, etc.)
 - StableLM(`stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc.)
+- Starcoder2(`bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc.)
 - Yi (`01-ai/Yi-6B`, `01-ai/Yi-34B`, etc.)
 
 Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):

csrc/ops.h

@@ -86,6 +86,15 @@ torch::Tensor awq_dequantize(
     int split_k_iters,
     int thx,
     int thy);
+
+torch::Tensor marlin_gemm(
+    torch::Tensor& a, 
+    torch::Tensor& b_q_weight,
+    torch::Tensor& b_scales, 
+    torch::Tensor& workspace,
+    int64_t size_m, 
+    int64_t size_n, 
+    int64_t size_k);
 #endif
 
 void squeezellm_gemm(
@@ -100,11 +109,13 @@ torch::Tensor gptq_gemm(
   torch::Tensor b_gptq_qzeros,
   torch::Tensor b_gptq_scales,
   torch::Tensor b_g_idx,
-  bool use_exllama);
+  bool use_exllama,
+  int bit);
 
 void gptq_shuffle(
   torch::Tensor q_weight,
-  torch::Tensor q_perm);
+  torch::Tensor q_perm,
+  int bit);
 
 void moe_align_block_size(
   torch::Tensor topk_ids,

csrc/pybind.cpp

@@ -52,11 +52,13 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     &rotary_embedding,
     "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
 
-  // Quantization ops
+// Quantization ops
 #ifndef USE_ROCM
   ops.def("awq_gemm", &awq_gemm, "Quantized GEMM for AWQ");
+  ops.def("marlin_gemm", &marlin_gemm, "Marlin Optimized Quantized GEMM for GPTQ");
   ops.def("awq_dequantize", &awq_dequantize, "Dequantization for AWQ");
 #endif
+ 
   ops.def("gptq_gemm", &gptq_gemm, "Quantized GEMM for GPTQ");
   ops.def("gptq_shuffle", &gptq_shuffle, "Post processing for GPTQ");
   ops.def("squeezellm_gemm", &squeezellm_gemm, "Quantized GEMM for SqueezeLLM");

csrc/quantization/gptq/matrix_view.cuh

@@ -146,6 +146,129 @@ public:
     __device__ __forceinline__ const uint32_t* item_uint32_ptr(int row, int column) { return &data[row / 8 * width + column]; }
 };
 
+class MatrixView_q2_row
+{
+public:
+    const uint32_t* data;
+    const int height;
+    const int width;
+
+    __device__ __forceinline__ MatrixView_q2_row(const uint32_t* data, const int height, const int width)
+        : data(data), height(height), width(width)
+    { }
+
+    __device__ __forceinline__ int item(int row, int column) const
+    {
+        int shift = (column & 0x0f) * 2;
+        return (data[row * width / 16 + column / 16] >> shift) & 0x03;
+    }
+
+    __device__ __forceinline__ void item2(int (&items)[2], int row, int column) const
+    {
+        int shift = (column & 0x0f) * 2;
+        uint32_t d = data[row * width / 16 + column / 16] >> shift;
+        items[0] = d & 0x03;
+        items[1] = (d >> 2) & 0x03;
+    }
+
+    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
+    {
+        int shift = (column & 0x0f) * 2;
+        uint32_t d = data[row * width / 16 + column / 16] >> shift;
+        items[0] = d & 0x03;
+        items[1] = (d >> 2) & 0x03;
+        items[2] = (d >> 4) & 0x03;
+        items[3] = (d >> 6) & 0x03;
+    }
+};
+
+class MatrixView_q3_row
+{
+public:
+    const uint32_t* data;
+    const int height;
+    const int width;
+
+    __device__ __forceinline__ MatrixView_q3_row(const uint32_t* data, const int height, const int width)
+        : data(data), height(height), width(width)
+    { }
+
+    __device__ __forceinline__ int item(int row, int column) const
+    {
+        int z_w = column * 3 / 32;
+        int z_mod =  column & 0x1f;
+
+        if (z_mod == 10) {
+            return (data[row * width * 3 / 32 + z_w] >> 30) | ((data[row * width * 3 / 32 + (z_w + 1)] << 2) & 0x4);
+        } else if (z_mod == 21) {
+            return (data[row * width * 3 / 32 + z_w] >> 31) | ((data[row * width * 3 / 32 + (z_w + 1)] << 1) & 0x6);
+        } else if (z_mod < 10) {
+            return (data[row * width * 3 / 32 + z_w] >> (z_mod * 3)) & 0x07;
+        } else if (z_mod < 21) {
+            return (data[row * width * 3 / 32 + z_w] >> (z_mod * 3  - 32)) & 0x07;
+        } else {
+            return (data[row * width * 3 / 32 + z_w] >> (z_mod * 3  - 64)) & 0x07;
+        }
+    }
+
+    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
+    {
+        int shift = (column & 0x1f);
+        uint32_t d;
+        if (shift <= 4) {
+            d = data[row * width / 32 * 3 + column * 3 / 32] >> (shift * 3);
+        } else if (shift == 8) {
+            d = (data[row * width / 32 * 3 + column * 3 / 32] >> 24) | ((data[row * width / 32 * 3 + column * 3 / 32 + 1] & 0x0f) << 8);
+        } else if (shift <= 16) {
+            d = data[row * width / 32 * 3 + column * 3 / 32] >> (shift * 3 - 32);
+        } else if (shift == 20) {
+            d = (data[row * width / 32 * 3 + column * 3 / 32] >> 28) | ((data[row * width / 32 * 3 + column * 3 / 32 + 1] & 0xff) << 4);
+        } else {
+            d = data[row * width / 32 * 3 + column * 3 / 32] >> (shift * 3 - 64);
+        }
+        items[0] = d & 0x07;
+        items[1] = (d >> 3) & 0x07;
+        items[2] = (d >> 6) & 0x07;
+        items[3] = (d >> 9) & 0x07;
+    }
+};
+
+class MatrixView_q8_row
+{
+public:
+    const uint32_t* data;
+    const int height;
+    const int width;
+
+    __device__ __forceinline__ MatrixView_q8_row(const uint32_t* data, const int height, const int width)
+        : data(data), height(height), width(width)
+    { }
+
+    __device__ __forceinline__ int item(int row, int column) const
+    {
+        int shift = (column & 0x03) * 8;
+        return (data[row * width / 4 + column / 4] >> shift) & 0xff;
+    }
+
+    __device__ __forceinline__ void item2(int (&items)[2], int row, int column) const
+    {
+        int shift = (column & 0x03) * 8;
+        uint32_t d = data[row * width / 4 + column / 4] >> shift;
+        items[0] = d & 0xff;
+        items[1] = (d >> 8) & 0xff;
+    }
+
+    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
+    {
+        int shift = (column & 0x03) * 2;
+        uint32_t d = data[row * width / 4 + column / 4] >> shift;
+        items[0] = d & 0xff;
+        items[1] = (d >> 8) & 0xff;
+        items[2] = (d >> 16) & 0xff;
+        items[3] = (d >> 24) & 0xff;
+    }
+};
+
 }  // namespace gptq
 }  // namespace vllm
 #endif