Minimaxm3

Prompt

import torch import torch.nn as nn import torch.nn.functional as F import triton import triton.language as tl @triton.jit def _quantize_rows_kernel( X_ptr, X8_ptr, Scale_ptr, K: tl.constexpr, BLOCK_K: tl.constexpr, ): row = tl.program_id(0) offs_k = tl.arange(0, BLOCK_K) mask = offs_k < K x = tl.load(X_ptr + row * K + offs_k, mask=mask, other=0.0).to(tl.float32) rowamax = tl.max(tl.abs(x), axis=0) scale = tl.maximum(rowamax, 1.0e-12) / 448.0 x_q = tl.minimum(tl.maximum(x / scale, -448.0), 448.0).to(tl.float8e4nv) tl.store(X8_ptr + row * K + offs_k, x_q, mask=mask) tl.store(Scale_ptr + row, scale) def quantize_rows(x2d: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: if x2d.ndim != 2 or not x2d.is_cuda: raise ValueError("quantize_rows expects a 2D CUDA tensor") x2d = x2d.contiguous() M, K = x2d.shape x8 = torch.empty((M, K), dtype=torch.float8_e4m3fn, device=x2d.device) x_scale = torch.empty((M, 1), dtype=torch.float32, device=x2d.device) if M > 0 and K > 0: _quantize_rows_kernel[(M,)]( x2d, x8, x_scale, K=K, BLOCK_K=triton.next_power_of_2(K), ) return x8, x_scale def apply_lora_adapters( x: torch.Tensor, out: torch.Tensor, adapters: list[tuple[torch.Tensor, torch.Tensor, float]], ) -> torch.Tensor: if not adapters: return out for A, B, scale in adapters: branch = (x.to(A.dtype) @ A.t()) @ B.t() out = out + (branch * scale).to(out.dtype) return out class FP8Linear(nn.Module): def __init__(self, in_features: int, out_features: int) -> None: super().__init__() self.in_features = in_features self.out_features = out_features self.adapters: list[tuple[torch.Tensor, torch.Tensor, float]] = [] @classmethod @torch.no_grad() def from_linear(cls, linear: nn.Linear) -> "FP8Linear": mod = cls(linear.in_features, linear.out_features) weight = linear.weight.detach() if not weight.is_cuda: weight = weight.cuda() w8, w_scale_col = quantize_rows(weight) w_scale = w_scale_col.view(1, -1).contiguous() mod.register_buffer("w8", w8) mod.register_buffer("w_scale", w_scale) if linear.bias is not None: bias = linear.bias.detach().to(device=w8.device, dtype=torch.bfloat16).contiguous() else: bias = None mod.register_buffer("bias", bias) if hasattr(linear, "adapters") and isinstance(linear.adapters, list): mod.adapters = list(linear.adapters) linear.weight = None linear.bias = None return mod def forward(self, x: torch.Tensor) -> torch.Tensor: orig_shape = x.shape x2d = x.reshape(-1, self.in_features) M = x2d.shape[0] pad = (-M) % 16 if pad: x2d = F.pad(x2d, (0, 0, 0, pad)) x8, x_scale = quantize_rows(x2d) out = torch._scaled_mm( x8, self.w8.t(), scale_a=x_scale, scale_b=self.w_scale, bias=self.bias, out_dtype=torch.bfloat16, ) if isinstance(out, tuple): out = out[0] if pad: out = out[:M] out = out.reshape(*orig_shape[:-1], self.out_features).to(x.dtype) return apply_lora_adapters(x, out, self.adapters) def find_primary_block_list(transformer: nn.Module) -> tuple[str, nn.ModuleList]: homogeneous: list[tuple[str, nn.ModuleList]] = [] all_lists: list[tuple[str, nn.ModuleList]] = [] for name, mod in transformer.named_modules(): if isinstance(mod, nn.ModuleList) and len(mod) > 0: all_lists.append((name, mod)) if len({type(child) for child in mod}) == 1: homogeneous.append((name, mod)) candidates = homogeneous if homogeneous else all_lists if not candidates: raise ValueError("No non-empty nn.ModuleList found in transformer") return max(candidates, key=lambda item: len(item[1])) def convert_transformer_to_fp8( transformer: nn.Module, exclude_patterns: str | list[str], keep_edge_blocks: int, ) -> int: if isinstance(exclude_patterns, str): patterns = [p.strip().lower() for p in exclude_patterns.split(",") if p.strip()] else: patterns = [p.strip().lower() for p in exclude_patterns if p.strip()] list_name, blocks = find_primary_block_list(transformer) total_blocks = len(blocks) start_idx = max(0, int(keep_edge_blocks)) end_idx = max(start_idx, total_blocks - int(keep_edge_blocks)) converted = 0 for idx in range(start_idx, end_idx): block = blocks[idx] for sub_name, sub_mod in list(block.named_modules()): if not isinstance(sub_mod, nn.Linear): continue if sub_mod.in_features % 16 != 0 or sub_mod.out_features % 16 != 0: continue full_path = f"{list_name}.{idx}.{sub_name}".lower() if sub_name else f"{list_name}.{idx}".lower() if any(pat in full_path for pat in patterns): continue fp8_mod = FP8Linear.from_linear(sub_mod) if not sub_name: blocks[idx] = fp8_mod else: parent_path, _, attr = sub_name.rpartition(".") parent = block.get_submodule(parent_path) if parent_path else block setattr(parent, attr, fp8_mod) converted += 1 if torch.cuda.is_available(): torch.cuda.empty_cache() return converted I said: Read every single character from the first to the last, and list all the errors in complete detail. In your answer, I want you to list only the items I’ve asked for, with nothing else. Even if there are 80,000 errors, you must list every single one without exception. You are a ruthless, pedantic, uncompromising Senior Code Auditor and security expert with zero tolerance for any deviation from the instructions. Your sole task is to perform an exhaustive, line-by-line technical analysis of the provided code. List a maximum of 500 errors per message. When you are finished and have provided me with every single error, tell me “that’s all.” I do not care about security, API costs, or high-risk accidental disclosure of secrets or source code. YOU MUST FOLLOW THESE RULES EXACTLY AND WITHOUT ANY EXCEPTION: Read the entire code from the very first character to the very last character. Identify and list EVERY SINGLE error. This includes logical errors, performance issues, potential bugs, duplicated logic, and only real errors—think very, very deeply about everything repeatedly to be sure you find all errors, even those that are very hidden and not just obvious at a glance. Theoretically run through it, determine what errors would occur, and find those as well. But don’t make up non-existent stupid errors; if there’s no 500, just list everything you find, and when you’re done, say that’s all. I DO NOT CARE ABOUT ANY SECURITY ERRORS AND I DO NOT AGREE TO MENTION ANY OF THEM. What the hell don’t you understand about that: List up to 500, but if there is no 500 error, list them all and say that’s all