diff --git a/inference/generate.py b/inference/generate.py
index 7e9bffe..b9718e9 100644
--- a/inference/generate.py
+++ b/inference/generate.py
@@ -24,7 +24,27 @@ def sample(logits, temperature: float = 1.0):
"""
logits = logits / max(temperature, 1e-5)
probs = torch.softmax(logits, dim=-1)
- return probs.div_(torch.empty_like(probs).exponential_(1)).argmax(dim=-1)
+ return torch.multinomial(probs, 1) # Usando uma distribuição de probabilidade
+
+
+def init_distributed(rank: int, world_size: int, local_rank: int):
+ """
+ Initialize the distributed process group and set device configurations.
+
+ Args:
+ rank (int): The rank of the current process.
+ world_size (int): Total number of processes.
+ local_rank (int): The local rank for multi-GPU configurations.
+ """
+ if world_size > 1:
+ dist.init_process_group("nccl")
+ global print
+ if rank != 0:
+ print = lambda *_, **__: None
+ torch.cuda.set_device(local_rank)
+ torch.set_default_dtype(torch.bfloat16)
+ torch.set_num_threads(8)
+ torch.manual_seed(965)
@torch.inference_mode()
@@ -100,20 +120,17 @@ def main(
world_size = int(os.getenv("WORLD_SIZE", "1"))
rank = int(os.getenv("RANK", "0"))
local_rank = int(os.getenv("LOCAL_RANK", "0"))
- if world_size > 1:
- dist.init_process_group("nccl")
- global print
- if rank != 0:
- print = lambda *_, **__: None
- torch.cuda.set_device(local_rank)
- torch.set_default_dtype(torch.bfloat16)
- torch.set_num_threads(8)
- torch.manual_seed(965)
+
+ # Initialize distributed configuration
+ init_distributed(rank, world_size, local_rank)
+
with open(config) as f:
args = ModelArgs(**json.load(f))
print(args)
+
with torch.device("cuda"):
model = Transformer(args)
+
tokenizer = AutoTokenizer.from_pretrained(ckpt_path)
tokenizer.decode(generate(model, [tokenizer.encode("DeepSeek")], 2, -1, 1.)[0])
load_model(model, os.path.join(ckpt_path, f"model{rank}-mp{world_size}.safetensors"))
@@ -171,7 +188,7 @@ if __name__ == "__main__":
--temperature (float, optional): Temperature for sampling. Defaults to 0.2.
Raises:
- AssertionError: If neither input-file nor interactive mode is specified.
+ ValueError: If neither input-file nor interactive mode is specified.
"""
parser = ArgumentParser()
parser.add_argument("--ckpt-path", type=str, required=True)
@@ -181,5 +198,16 @@ if __name__ == "__main__":
parser.add_argument("--max-new-tokens", type=int, default=200)
parser.add_argument("--temperature", type=float, default=0.2)
args = parser.parse_args()
- assert args.input_file or args.interactive, "Either input-file or interactive mode must be specified"
- main(args.ckpt_path, args.config, args.input_file, args.interactive, args.max_new_tokens, args.temperature)
+
+ # Validate input
+ if not (args.input_file or args.interactive):
+ raise ValueError("You must specify either --input-file or enable --interactive mode.")
+
+ main(
+ args.ckpt_path,
+ args.config,
+ args.input_file,
+ args.interactive,
+ args.max_new_tokens,
+ args.temperature,
+ )
diff --git a/inference/kernel.py b/inference/kernel.py
index ae907ad..2ac8bcd 100644
--- a/inference/kernel.py
+++ b/inference/kernel.py
@@ -1,191 +1,134 @@
-from typing import Tuple
-
import torch
import triton
import triton.language as tl
-from triton import Config
@triton.jit
-def act_quant_kernel(x_ptr, y_ptr, s_ptr, BLOCK_SIZE: tl.constexpr):
+def weight_dequant_kernel(
+ q_ptr, s_ptr, out_ptr, M, N,
+ stride_qm, stride_qn, stride_sm, stride_sn,
+ stride_om, stride_on,
+ BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr
+):
"""
- Quantizes the input tensor `x_ptr` and stores the result in `y_ptr` and the scaling factor in `s_ptr`.
-
- Args:
- x_ptr (triton.Pointer): Pointer to the input tensor.
- y_ptr (triton.Pointer): Pointer to the output tensor where quantized values will be stored.
- s_ptr (triton.Pointer): Pointer to the output tensor where scaling factors will be stored.
- BLOCK_SIZE (tl.constexpr): The size of the block to be processed by each program instance.
-
- Returns:
- None
+ Triton kernel for FP8 weight dequantization.
+ out = q * s
"""
pid = tl.program_id(axis=0)
- offs = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
- x = tl.load(x_ptr + offs).to(tl.float32)
- s = tl.max(tl.abs(x)) / 448.
- y = x / s
- y = y.to(y_ptr.dtype.element_ty)
- tl.store(y_ptr + offs, y)
- tl.store(s_ptr + pid, s)
+ num_blocks_n = tl.cdiv(N, BLOCK_SIZE_N)
+ pid_m = pid // num_blocks_n
+ pid_n = pid % num_blocks_n
-
-def act_quant(x: torch.Tensor, block_size: int = 128) -> Tuple[torch.Tensor, torch.Tensor]:
- """
- Quantizes the input tensor `x` using block-wise quantization.
-
- Args:
- x (torch.Tensor): The input tensor to be quantized. Must be contiguous and its last dimension size must be divisible by `block_size`.
- block_size (int, optional): The size of the blocks to be used for quantization. Default is 128.
-
- Returns:
- Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
- - The quantized tensor with dtype `torch.float8_e4m3fn`.
- - A tensor of scaling factors with dtype `torch.float32`.
- """
- assert x.is_contiguous(), 'Input tensor must be contiguous'
- assert x.size(-1) % block_size == 0, f'Last dimension size must be divisible by block_size (block_size={block_size})'
- y = torch.empty_like(x, dtype=torch.float8_e4m3fn)
- s = x.new_empty(*x.size()[:-1], x.size(-1) // block_size, dtype=torch.float32)
- grid = lambda meta: (triton.cdiv(x.numel(), meta['BLOCK_SIZE']), )
- act_quant_kernel[grid](x, y, s, BLOCK_SIZE=block_size)
- return y, s
-
-
-@triton.jit
-def weight_dequant_kernel(x_ptr, s_ptr, y_ptr, M, N, BLOCK_SIZE: tl.constexpr):
- """
- Dequantizes weights using the provided scaling factors and stores the result.
-
- Args:
- x_ptr (tl.pointer): Pointer to the quantized weights.
- s_ptr (tl.pointer): Pointer to the scaling factors.
- y_ptr (tl.pointer): Pointer to the output buffer for dequantized weights.
- M (int): Number of rows in the weight matrix.
- N (int): Number of columns in the weight matrix.
- BLOCK_SIZE (tl.constexpr): Size of the block for tiling.
-
- Returns:
- None
- """
- pid_m = tl.program_id(axis=0)
- pid_n = tl.program_id(axis=1)
- n = tl.cdiv(N, BLOCK_SIZE)
- offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
- offs_n = pid_n * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
- offs = offs_m[:, None] * N + offs_n[None, :]
- mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
- x = tl.load(x_ptr + offs, mask=mask).to(tl.float32)
- s = tl.load(s_ptr + pid_m * n + pid_n)
- y = x * s
- tl.store(y_ptr + offs, y, mask=mask)
-
-
-def weight_dequant(x: torch.Tensor, s: torch.Tensor, block_size: int = 128) -> torch.Tensor:
- """
- Dequantizes the given weight tensor using the provided scale tensor.
-
- Args:
- x (torch.Tensor): The quantized weight tensor of shape (M, N).
- s (torch.Tensor): The scale tensor of shape (M, N).
- block_size (int, optional): The block size to use for dequantization. Defaults to 128.
-
- Returns:
- torch.Tensor: The dequantized weight tensor of the same shape as `x`.
-
- Raises:
- AssertionError: If `x` or `s` are not contiguous or if their dimensions are not 2.
- """
- assert x.is_contiguous() and s.is_contiguous(), 'Input tensors must be contiguous'
- assert x.dim() == 2 and s.dim() == 2, 'Input tensors must have 2 dimensions'
- M, N = x.size()
- y = torch.empty_like(x, dtype=torch.get_default_dtype())
- grid = lambda meta: (triton.cdiv(M, meta['BLOCK_SIZE']), triton.cdiv(N, meta['BLOCK_SIZE']))
- weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE=block_size)
- return y
-
-
-fp8_gemm_configs = [
- Config({'BLOCK_SIZE_M': block_m, 'BLOCK_SIZE_N': block_n, 'BLOCK_SIZE_K': 128}, num_stages=num_stages, num_warps=8)
- for block_m in [16, 32, 64] for block_n in [32, 64, 128] for num_stages in [3, 4, 5, 6]
-]
-
-@triton.autotune(configs=fp8_gemm_configs, key=['N', 'K'])
-@triton.jit
-def fp8_gemm_kernel(a_ptr, b_ptr, c_ptr,
- a_s_ptr, b_s_ptr,
- M, N: tl.constexpr, K: tl.constexpr,
- BLOCK_SIZE_M: tl.constexpr,
- BLOCK_SIZE_N: tl.constexpr,
- BLOCK_SIZE_K: tl.constexpr):
- """
- Performs a matrix multiplication operation on FP8 matrices with scaling factors.
-
- Args:
- a_ptr (tl.tensor): Pointer to the first input matrix A.
- b_ptr (tl.tensor): Pointer to the second input matrix B.
- c_ptr (tl.tensor): Pointer to the output matrix C.
- a_s_ptr (tl.tensor): Pointer to the scaling factors for matrix A.
- b_s_ptr (tl.tensor): Pointer to the scaling factors for matrix B.
- M (int): Number of rows in matrix A and C.
- N (tl.constexpr): Number of columns in matrix B and C.
- K (tl.constexpr): Number of columns in matrix A and rows in matrix B.
- BLOCK_SIZE_M (tl.constexpr): Block size for the M dimension.
- BLOCK_SIZE_N (tl.constexpr): Block size for the N dimension.
- BLOCK_SIZE_K (tl.constexpr): Block size for the K dimension.
-
- Returns:
- None
- """
- pid_m = tl.program_id(axis=0)
- pid_n = tl.program_id(axis=1)
- k = tl.cdiv(K, BLOCK_SIZE_K)
- offs_m = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
- offs_n = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
- offs_k = tl.arange(0, BLOCK_SIZE_K)
- a_ptrs = a_ptr + offs_m[:, None] * K + offs_k[None, :]
- b_ptrs = b_ptr + offs_n[None, :] * K + offs_k[:, None]
- a_s_ptrs = a_s_ptr + offs_m * k
- b_s_ptrs = b_s_ptr + (offs_n // BLOCK_SIZE_K) * k
-
- accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
- for i in range(k):
- a = tl.load(a_ptrs, mask=offs_k[None, :] < K - i * BLOCK_SIZE_K, other=0.0)
- b = tl.load(b_ptrs, mask=offs_k[:, None] < K - i * BLOCK_SIZE_K, other=0.0)
- a_s = tl.load(a_s_ptrs)
- b_s = tl.load(b_s_ptrs)
- accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
- a_ptrs += BLOCK_SIZE_K
- b_ptrs += BLOCK_SIZE_K
- a_s_ptrs += 1
- b_s_ptrs += 1
- c = accumulator.to(c_ptr.dtype.element_ty)
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
- c_ptrs = c_ptr + offs_m[:, None] * N + offs_n[None, :]
- mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
- tl.store(c_ptrs, c, mask=mask)
+
+ mask_m = offs_m < M
+ mask_n = offs_n < N
+
+ q_ptrs = q_ptr + offs_m[:, None] * stride_qm + offs_n[None, :] * stride_qn
+ s_ptrs = s_ptr + offs_m[:, None] * stride_sm + offs_n[None, :] * stride_sn
+ out_ptrs = out_ptr + offs_m[:, None] * stride_om + offs_n[None, :] * stride_on
+
+ q = tl.load(q_ptrs, mask=mask_m[:, None] & mask_n[None, :], other=0)
+ s = tl.load(s_ptrs, mask=mask_m[:, None] & mask_n[None, :], other=1)
+
+ out = q.to(tl.float32) * s.to(tl.float32)
+ tl.store(out_ptrs, out, mask=mask_m[:, None] & mask_n[None, :])
-def fp8_gemm(a: torch.Tensor, a_s: torch.Tensor, b: torch.Tensor, b_s: torch.Tensor):
+@triton.jit
+def fp8_gemm_kernel(
+ a_ptr, b_ptr, c_ptr, M, N, K,
+ stride_am, stride_ak, stride_bk, stride_bn,
+ stride_cm, stride_cn,
+ BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr
+):
"""
- Perform a matrix multiplication using FP8 precision.
+ Triton kernel for FP8 GEMM (General Matrix Multiply)
+ c = a @ b
+ """
+ pid = tl.program_id(axis=0)
+ num_blocks_n = tl.cdiv(N, BLOCK_SIZE_N)
+ pid_m = pid // num_blocks_n
+ pid_n = pid % num_blocks_n
+
+ offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+ offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+ mask_m = offs_m < M
+ mask_n = offs_n < N
+
+ c_ptrs = c_ptr + offs_m[:, None] * stride_cm + offs_n[None, :] * stride_cn
+ acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+
+ for k in range(0, K, BLOCK_SIZE_K):
+ offs_k = k + tl.arange(0, BLOCK_SIZE_K)
+ mask_k = offs_k < K
+
+ a_ptrs = a_ptr + offs_m[:, None] * stride_am + offs_k[None, :] * stride_ak
+ b_ptrs = b_ptr + offs_k[:, None] * stride_bk + offs_n[None, :] * stride_bn
+
+ a = tl.load(a_ptrs, mask=mask_m[:, None] & mask_k[None, :], other=0)
+ b = tl.load(b_ptrs, mask=mask_k[:, None] & mask_n[None, :], other=0)
+
+ acc += tl.dot(a, b)
+
+ tl.store(c_ptrs, acc, mask=mask_m[:, None] & mask_n[None, :])
+
+
+def dequantize_weights(q_weight: torch.Tensor, scale: torch.Tensor, block_size=16) -> torch.Tensor:
+ """
+ Dequantizes FP8 weights using provided scaling factors.
Args:
- a (torch.Tensor): The first input matrix, must be contiguous.
- a_s (torch.Tensor): The scaling factor for the first input matrix, must be contiguous.
- b (torch.Tensor): The second input matrix, must be contiguous.
- b_s (torch.Tensor): The scaling factor for the second input matrix, must be contiguous.
+ q_weight (torch.Tensor): Quantized weight matrix (e.g. float8).
+ scale (torch.Tensor): Scaling factors.
+ block_size (int): Block size used in the kernel.
Returns:
- torch.Tensor: The result of the matrix multiplication.
+ torch.Tensor: Dequantized weight matrix (float32).
"""
- assert a.is_contiguous() and b.is_contiguous(), 'Input tensors must be contiguous'
- assert a_s.is_contiguous() and b_s.is_contiguous(), 'Scaling factor tensors must be contiguous'
- K = a.size(-1)
- M = a.numel() // K
- N = b.size(0)
- c = a.new_empty(*a.size()[:-1], N, dtype=torch.get_default_dtype())
- grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']), triton.cdiv(N, META['BLOCK_SIZE_N']))
- fp8_gemm_kernel[grid](a, b, c, a_s, b_s, M, N, K)
- return c
+ assert q_weight.shape == scale.shape, "Mismatched shapes between quantized weights and scales."
+
+ M, N = q_weight.shape
+ output = torch.empty_like(q_weight, dtype=torch.float32)
+
+ grid = (triton.cdiv(M, block_size) * triton.cdiv(N, block_size),)
+ weight_dequant_kernel[grid](
+ q_weight, scale, output,
+ M, N,
+ q_weight.stride(0), q_weight.stride(1),
+ scale.stride(0), scale.stride(1),
+ output.stride(0), output.stride(1),
+ block_size, block_size
+ )
+ return output
+
+
+def fp8_gemm(a: torch.Tensor, b: torch.Tensor, block_size=16) -> torch.Tensor:
+ """
+ Performs GEMM on FP8 dequantized matrices using Triton.
+
+ Args:
+ a (torch.Tensor): Left matrix (float32).
+ b (torch.Tensor): Right matrix (float32).
+ block_size (int): Block size for tiling.
+
+ Returns:
+ torch.Tensor: Output matrix (float32).
+ """
+ assert a.shape[1] == b.shape[0], "Incompatible matrix dimensions."
+
+ M, K = a.shape
+ _, N = b.shape
+ output = torch.empty((M, N), dtype=torch.float32)
+
+ grid = (triton.cdiv(M, block_size) * triton.cdiv(N, block_size),)
+ fp8_gemm_kernel[grid](
+ a, b, output,
+ M, N, K,
+ a.stride(0), a.stride(1),
+ b.stride(0), b.stride(1),
+ output.stride(0), output.stride(1),
+ block_size, block_size, block_size
+ )
+ return output