diff --git a/config.json b/config.json index 17b3c78..c9ceb67 100644 --- a/config.json +++ b/config.json @@ -10,16 +10,16 @@ }, "bos_token_id": 130004, "eos_token_id": 130005, + "mask_token_id": 130000, "gmask_token_id": 130001, + "pad_token_id": 3, "hidden_size": 4096, "inner_hidden_size": 16384, "layernorm_epsilon": 1e-05, - "mask_token_id": 130000, "max_sequence_length": 2048, "model_type": "chatglm", "num_attention_heads": 32, "num_layers": 28, - "pad_token_id": 3, "position_encoding_2d": true, "quantization_bit": 4, "quantization_embeddings": false, diff --git a/modeling_chatglm.py b/modeling_chatglm.py index 52ac051..760eda3 100644 --- a/modeling_chatglm.py +++ b/modeling_chatglm.py @@ -918,7 +918,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel): elif input_ids is not None: batch_size, seq_length = input_ids.shape[:2] elif inputs_embeds is not None: - batch_size, seq_length, _ = inputs_embeds.shape[:2] + batch_size, seq_length = inputs_embeds.shape[:2] else: raise ValueError("You have to specify either input_ids or inputs_embeds") @@ -972,9 +972,8 @@ class ChatGLMModel(ChatGLMPreTrainedModel): if attention_mask is None: attention_mask = torch.zeros(1, 1, device=input_ids.device).bool() - else: - attention_mask = attention_mask.to(input_ids.device) + attention_mask = attention_mask.to(hidden_states.device) for i, layer in enumerate(self.layers): diff --git a/quantization.py b/quantization.py index 5be8b0b..49b53de 100644 --- a/quantization.py +++ b/quantization.py @@ -17,6 +17,7 @@ logger = logging.get_logger(__name__) try: from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up + class Kernel: def __init__(self, code: bytes, function_names: List[str]): self.code = code @@ -26,6 +27,7 @@ try: for name in self._function_names: setattr(self, name, KernelFunction(self._cmodule, name)) + quantization_code = "$QlpoOTFBWSZTWU9yuJUAQHN//////////f/n/8/n///n//bt4dTidcVx8X3V9FV/92/v4B7/AD5FBQFAAAChSgKpFCFAFVSigUAAAEKhSgUUqgFBKigqVREQAABQBQIANDTTIGI00BkZBkNGE0A0BkBkGQGRkaNAaAGQNBoGgDIAAYIGTI0DQAQAaGmmQMRpoDIyDIaMJoBoDIDIMgMjI0aA0AMgaDQNAGQAAwQMmRoGgAgA0NNMgYjTQGRkGQ0YTQDQGQGQZAZGRo0BoAZA0GgaAMgABggZMjQNABABoaaZAxGmgMjIMhowmgGgMgMgyAyMjRoDQAyBoNA0AZAADBAyZGgaAAmqU1NEgJqnptU/Sn4jRR6J6epk2pqb1Q/SgAPUGgyNNGjQ2SBpoAZAAGg0NB6mgDIAAAAA2oaApSREBNAARhGiYEaEwU8pvImlP0k2aam1GaGqbFNM1MHpTwmkepmyU9R6nqPKekHqNNPUxNGhp6n6p6QaZ6o9TG1GMqcoV9ly6nRanHlq6zPNbnGZNi6HSug+2nPiZ13XcnFYZW+45W11CumhzYhchOJ2GLLV1OBjBjGf4TptOddTSOcVxhqYZMYwZXZZY00zI1paX5X9J+b+f4e+x43RXSxXPOdquiGpduatGyXneN696M9t4HU2eR5XX/kPhP261NTx3JO1Ow7LyuDmeo9a7d351T1ZxnvnrvYnrXv/hXxPCeuYx2XsNmO003eg9J3Z6U7b23meJ4ri01OdzTk9BNO96brz+qT5nuvvH3ds/G+m/JcG/F2XYuhXlvO+jP7U3XgrzPN/lr8Sf1n6j4j7jZs+s/T0tNaNNYzTs12rxjwztHlnire3Nzc3N1wuBwOBwXBvZfoHpD7rFmR99V5vj3aXza3xdBbXMalubTg/jIv5dfAi54Pdc75j4z412n3Npj3Ld/ENm7a3b/Cod6h/ret1/5vn/C+l+gdslMvgPSLJ8d8q+U66fevYn/tW1chleEtNTGlcHCbLRlq0tHzF5tsbbZZfHjjLgZu42XCuC3NrdjTasZGNzgxPIrGqp7r3p7L2p5XjnpPSmTd5XtzqnB6U87zzg1Ol0zd0zsLszxR6lkxp35u6/teL0L0W922cR7Lu1lpL9CsHirzuM2T+BgsyViT6LHcm0/Vr6U/7LGGyJeqTEjt0PHWhF5mCT7R9mtlDwriYv0Tyr/OxYt6qp5r0mPVT0608TqnqMZaarU2nFwrTzzlrs1ed7z1ux60wyr4ydCaTi3enW8x68x0zU7tXSlcmPSW1mGpWJMg4zmPC2lK96tp0OE80y4MfEvnZj8zGluR6b22ki1Ou9V2nCd9xovcPvcYMZYy0lvN60ScZ45vN6yeCeeXFb1lVjnnCar5fwXwE2bzJ4HI1XVPXfXZMm44GUsMpYsmLB65TuVdm0cl0b+i/wGNN66XjeV7zuPpHcnK/juhhjdfId5jMdE5nN0dGmmm2zZs2cexD5n9p/dY352XsvXHaZNWWsmmS1atjR452nYudzvqv2HMRyvNNnlMcDl3R2+yx2uVrBubTW9icHDVtbNXlZm7jma1rM4VurZZd2y6nUau7ZXZ7bVU+mnoOVxZGMrVmvX60605JwmzGZhhhjTWtaaaMaaGTGmNMZasY0iX8VMUl8eepaIrzGSpemWOQyZORk2bNpjUybMmxqYmknCGCFynutfksaZpjTNMaaatM0xsxcGR0sociNqxNSmhhR1ZJPbsn8qyF0t2qH6iYBclclalbtTTcHTDsPaX6rlnElph2Jyumumtynv2Kk8GI7rsvXbIcJgHJOSaSXnnGaI3m87RtVXJOZ/YtgdTE6Wpha6ZlE8ayXkef1fh602r2WwvfMXtMdLlkfnLFdYYwYso+bWqm7yJqHXZGw2nrS5ZanSYnWlxBxMF1V940K2wdrI7R6OYf7DGGamMmTSbRhlS45xmVOumF1EyPCmHrrN8wwZOOrdNtLeMtzFzDlWnfTBxMk2NaXIZHBYxYLD4w8yju0ao65Vz1OIXoS9dLanwCe1PWrYuWMqf1if1z2k2yYfKJ741PDgno1ZQ8DRqvUny3mNoWTzGO6m1DkrJI8JiR5cSd+vZdGOO8nrMoc5+NDUFsMSXaZJeNlMmGLtJsovOsUp7I9S5VojKxF6bTVEelXqlfJobQr3LozSh2Jk7VcrVMfhXqszGWMzNqGhqZY0OadxkyyMssKugZR0KNFXBHlqwmJgTE/BNVMk6ItJXZMR0H47GpXv/DMOvNkmVuaV1PRfEdxuqc7Hcd+ZV/zTLaRxWk0nl9CdCeM6mn5rstHIBcpiuwmUZXeq81DacHI2rmrZ5SuE5mOZd6LQrZg9mx32TprA8BMo5jKN6yLTCi3WzQaZSuhzTtM1fUTGVpG8Tw+KXI0tjEpiWxtLYynOlktSbVlaI5kxP8TDH8kx50xoxi5KcA4pcja8KWLRlO/Ks6q06ergnvm1ca3Tq8Uw7LTUsmWyctXPWmpitl/uvGcWTGXGuAXDfhqazGmjkxcJW5hMMMMpYsXl2TZYtVOddG3XCarUt6Ptq9CZXSNzyuRzqRZOjsxdBbFVz6OA5HI43r1jityVlVpVkxmOsyaYWE1NTGq1sOVh36mHMcxtSvcy70edG0ZGR3I1Go1GRlV7mWWo1G0ZGRqlvH40l7o4m5xMWLLLYyNjnqc8556mdPqLJ31n/1nWOncxzG1tizrHs/Z+d2vP/B/l8wdJ6rHUn2nbbDq4p6htFtYzMMMTaZis1K5GKzGNmxhmUx2DDlZ/qNnIx41xnaMfCZWYaZWtNLTNW8ND4Fw1MyZOCdM428suKG1ehW8TesOydg7J+YYcD4cYR+8dFK6M4E3HM9ZfRNNL+Sn6rsl4DsrDl2HpPCnfxjGXtbZtYys1ttlyJ4T+BvexjGWRjMszK4Jpc77D3GyuVD7q0+G8m9G+2+rGm7cOR2y7FdtY2XUYx/oNlfRYxhMYyYZkyyg55enna9Kt/FFi6GMMwYwdwxWgxGMLKYmUyGExTKMZkMFhkymKuh0NOBNnBu+23LdwDoZYYzGGMxtORaTU1pjTGWTTGGtMrNWUsyyTTLLG1qy2ZjbK2DBllWqxMtBMaYZQmcE7zvvRcTkclUwdkxTaSdyySt/7fpL+T1v516Ji97fwr5JbLu305zMn5+GMTTZ9F+y7ExwmGVfG44yxn3dLv6l5i+Wth1jCrDq21nW9LqvvDzz3Vf3LLH/O/32TJ/erx3bXftO4eF+G956D952K/An4NfvOpjFjExjevP/UmE0fIoZXx6/w6lX/no3D0bLt+ixjieBM6ksRd0yB4Lt2SwYNE+gd1detlZWUnpiZfGfFaK+4PyCa/v18V8X75pe9fLXzp7l3VjF76vWZmHwGz1IZNWT7b8yddJ4q5kyrVdfru6atWc7bVYztL9Jf4GXvT+Y8m9/YsXP6H018a8D4XVOqvfzqeR+6yZOD8dPv0+U7/q5Pl+2dNb0MjzGVH5p6MNQ7cOWvw62U9aHE8DprDek+McLyvDz+te+9Zhq5+YTruufMcWMabqysTmZVWjKPfnK0wyVcrsuhjZRdLkHNvD72b9abriOSGIxiLixMOoalNPXzy+wT/tf+U6HHONfsz+xe8ufHBdQWWGWLA9if0rsnmrxK5LvRZQeWsTCsrmOYy8VteVfuRfcVTtDLItLIsMYxZLdU/DbtSemxF6Z6Zo5WBXE4tFdCyVMMXMTEMZXVlS6Xec2T4e0tHsRcEuWshcJ2YsNF5rUx1E8ifCq6Z+ZP7qdCeu/aTwFd53l16/o0NOw6O3dLavP4Hbi4RdmuDk6DoYaninC0+o4uZjbJ7Rxeu0/FbuFg+q7DVS6fQe0rZ6NDGUNNU6DEqOaLTicKnYZMnBWruljQxoaS3dZhocDge0bSTyOvdAbG5hxe2xji7E/L55xX13wWNDi6HCekcFxfCPGxY0MXC+s7afWaMdDyjyr+o8Rudm/NabOZvdl274zH4f5XK9z6On1Pe/K5TdPAslg77BjuO6Y3eO7GqvOPG/stknp1leyvLL0Z7bl9I4noMvLkzytLhWYzrOZzLXCORe028rORzOg4N/L0HlMOQ3Pgmnbb6KczlabORpu980q37TBqRu0/p3PO6234Bl03Ynuz+9W7gnsEcmvYaYY3aMYY0wx3pYd+ujsXauWdaY5Xkbtl23fPzFHiDB/QMo0yFjBllYxTQYYyxkrwn7JufwJ/PfgJ+C83X69ni6zvXcnyXabv0ncbLwsceS+RNlyN2mnneJtX0ngYO0+e+0+UnA+Wch3ji8hj5an4h+i6XBySU4n+R0roVcbw5yvHrmr4Yw8Y7x6c+9POPYHI5HI5HI5HI5HGXGww4nE4nrVyOR8XeqPEO7PLOiukYa3Novk5hV4cdtYZLI93e+uxff2jRo0aNGjRo0aNG1bVtW1dy3m83m8+tQ5ZzHw3nObwOu8La9Rc1dtkdS8A3eTk823tnktXWlxN6Oixe06zrN70Isd9jiOgZFq9yfkPqP/SLhN2Myl8jDM43bl1nbcb4cO57jlh8Jow6pzXZdL4dyODTuuhu77FyO27DdwdRxmvO+O+3N2+BdqyTwLHVczDVY4UPE4O66/ZO2cx1LFzVdSXtF7G4HMbrauOHRw6c8FdZ5m9fHZHYZXfTlZquyynSyTTKke6vcffSD9pzPA/G7n7jxPmuhc1DHMynPMrGL6AdewYmwu5ko+UUyTwrMv27rPH1v1nGqd87+p6N6LU8k3NEng53xXyHS97+44OSg/sy/hn+Se6yfYNjW0/uTgP+PvWYzLMmjhcLB/gGpri6H83/84eUXWT6T9Hsv7785z/7z4icpW+zfXypuR7rx/gMdZb1/wC678pcs8/2a3mDitGHxl9mfPlll5MafWWqxk/eYuTDgcNMzDGWLWvsuglNxs53GtN6uWpktlW1tZZYcuinMMWmnNnJydze3b2Y1McBxrBkXw799izLMZZYyy0TkbsGM4p03S2uVu5s/XXUdSdec6smVxZYYGpVmT8A+8ajuEyV5FatkvVru2x6uxGXXbH4A+jvgP4GMYy3iPLXzq/6z65+E005ey+cwMZD3fZcqc6xpjTFjQ0P3U+e++cPYmTIwj0nrK5NPTfl3WvpfLtXDcb2HQMudYOxFXQBor4L4T6vrOauFctYXJQ++NUWmJe5bmx1jDiZS1dTqWxo4GR8jm3fttpmPHppk9PEyv4/y8/sO07XacOmcqc0x2Vi9BvNJvN5oW8x4mOsydpidRxMYJPx06m1bqPzq9KtK8sxXNXFodD/+MYYaJTLwOhc9brCsV18oOR1i4tXChyTkq4lf4y1Ke+9axjDHqs1mfBbMXuP4Hzi+X7t8vzv7bHerrUPgPCxhjre4fXdfLNtNM+Jd+Zdh8xd8wP87uNPoPgv4W7/5P2BuxfsMabNnMnza+54Pdi5U671GPZY8CehX8Voeoo7FHpkeEc6715FwHZrIrUrHaviPUbPZHND+IhczrP6FcYvhOZ0Di/ETt0OI+YwNWR9r7tpf6WDeZKZDB1+z2IthOl1mPyb5FluvEx9h9d0NnM0Y1XPFkWIsk1WotJ0PBMmkvjvQTd0e71tfeV+8r8lQ/tpzpsmxJ+InrI/dj2UajUajVTUajatRqNRtGo1Go1Go4wjeMpZFMVV9CHbofPraLsJ3JpWV2XOoanCuFky4y3PPNxucK2uKC1Lbdb1eo+m5XomN6HfeZsabHLHRX/K+offtNGGmHWctcVcG44MdSqsOLY9VzX+Zxfxn2HPdWTpzWvkrtJ8M5zorrKcquRytJ5N5DZmcaW02l76nWO+BqPXm1A2Ry/0q71dH/mqrqeFjkYxjEXtsX8qubTk67rGycyqsdm4tZx5D6D5hhi0waaWmiaMP81Yjii5qxPlPuU/GfTL1Y5E6Jyfiq63qTa39A4J0sOGDgO9WF9bOXl0XfPRbsY2bPNKPy1YrFYrFYmRhhlTIyMjJWJYZHXuCXI8OoXsvfljGLFicNifpp2XunoPiG1wtx3p1Tah+/DD66OnVtVXP9rKbVxOnL0tR/rHtqB5UDErUVcl11D4qqvjpOcxX7armUNJB3LpW6bxVvD08e8h3odKKvyCFZBdSh2FVcST9xV3n3T8t1j7Kr9qgrqXg+13Pt5U7JCvFXVIV1YG5lRhkVYZJYYDDD4KOIMoHCp26WS8GB7uBh2zIdgq/PKyInjV2STShuoapUdCpX1yTwqq/z1VvET7Kh5nVPkO8YyxjLt2MaaMmWTLQvx3qnzltnXW0p2jxgbEtSny/Osv8Y9pLMXYoHVPAhkVdWVeODhR6q9/Sxe2liwwZWMVvFXfRkeIDxAePUPIrdJ4ey6yquzH+PD/bUOWAu05qVHtFd8rrKHSoeNIOUqrYr3FXyToqfYJgwmJdKpXXOwYYegNNGMzfZPp/t3t/DVs4zjNTN61rRqaWaa4NYbRjTa0tWwy2Y2tGN8ZO8ofNKq4j9SL7I+cSm4/6ovLV5HNXLI0jJidwrtk6ynCaP6Z++GjRlWS3tLeW129Mi9evxU9mtz6s5J3Z7M2ngTgnKvmpomxpaLCzPfmx0JWE+m3NLDDGOX47RctdYYNK5jakdqLkRlI39n590T5zctGSwwZZDJj6kW8XSi6ot2MmWWJ0DUT3nuvebBudScjZ79g8cWJ8av0k+/bE5WKd5MdbFpbDVMxu1DVMmtNZGJvq1mtRbn6M+g/kP0FwDwr7quZs7xosNGpbscyxhhd9TyJyFwbLcxlTasg75vW7TsV5K7ji44XPMMrdoj+Y3rT0Hie62nlYV/pwczzOmdLqLhYkzGMzCZWGMQzGMSsZYY6Di1t4nlJ+Em63mJxrVLxPbYxNEdgc1dU2iOKyoYYWjNrEeHTYybVk0atSa7ehuwsWMWTqn1TrnS6hYsi71d1+s+k+ic70e20fzE/VaTdxT9ZtU4GIXdeNx3X77guYYfpHeTQjaMX6brOu4OY4K7Y2d9mbHarI5ox3p4GpJ2Vd/Tst60f7j999pppjR+Q/Qf8J/VaORs3cji7FfFuN61+ui9s8hix1OCh5KGVV23BPXvZfz3CLyHpix+exi8z/KnCnosY2eunor+cxyPO/xJ0vKey9OvE9VjqaYu0x3Z3jd6o2b1T12D+F8l232lwaaacD5LE8LBxu7WTlbWraWpew8Xexjel3E+wWD4APITdNqR8F3R3T0lunCQ4GaE9R37DxeCYfcHi4xci5ovKfxVs55y2hf+65E/Xdp6jR5nrebTmi5incpkyOjs50JvrZwstbbW6kfuuQw+2mykf/EXNFzxfKTrxew929TR6bWnGL//F3JFOFCQT3K4lQ" kernels = Kernel( @@ -68,7 +70,8 @@ class W8A16Linear(torch.autograd.Function): class W8A16LinearCPU(torch.autograd.Function): @staticmethod - def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width, quantization_cache=None): + def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width, + quantization_cache=None): ctx.inp_shape = inp.size() ctx.weight_bit_width = weight_bit_width out_features = quant_w.size(0) @@ -91,15 +94,17 @@ class W8A16LinearCPU(torch.autograd.Function): default_cpu_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "quantization_kernels.c") default_cpu_kernel_code = "QlpoOTFBWSZTWXLbSoQAAgzbgERwQXxmTwAAr/ff3kABt0Q2oRVT0hpo9RtEAAAAyBEiSQ9EGjQGQAAAwANGhowjJoNGmgMEUplMTNSMJ5TQaDJpsoMyRMj8P4mZzFSVVwqSXG8GG7MlVwiToYEQwVD7noBxMhNfkeZYtYFtbgOBUSIGtIQjhNHCEnPJsadhb3yBmRIOD3TeAtNLSaU5GgvKUBWSNuuOIHmVt0YhW6rsmDMDUjeUJGJ64R1Jm5lrh0Aa0tKjhFwPdWcGogxLDSXPWQUWTM8Sd3Qz1HMYNxx3HMeiNqNo4jeRDEfZ3gUSHIcU/heomq0vEzL1Msz5KKGxH8FrNOYw3KaxdqaEmNHYMxJFgQbR0DyRknL2L4kwUSxKRdhjRpEtUqilVfggFL1klaMS3PPRDfNqbBOPWO7m4JTVGhS9QTBDDJaEbLbrUQNB+IpJSKQbG5SZZ5gkwJEhJ3aYKJipZ/i7kinChIOW2lQg" -default_cpu_parallel_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "quantization_kernels_parallel.c") +default_cpu_parallel_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), + "quantization_kernels_parallel.c") default_cpu_parallel_kernel_code = "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" cpu_kernels = None class CPUKernel: - def __init__(self, kernel_file="", source_code=default_cpu_kernel_code_path, compile_parallel_kernel=None, parallel_num=None): - self.load =False + def __init__(self, kernel_file="", source_code=default_cpu_kernel_code_path, compile_parallel_kernel=None, + parallel_num=None): + self.load = False self.int8WeightExtractionFloat = None self.int4WeightExtractionFloat = None self.int4WeightCompression = None @@ -137,7 +142,8 @@ class CPUKernel: kernel_file = source_code[:-2] + ".so" if compile_parallel_kernel: - compile_command = "gcc -O3 -fPIC -pthread -fopenmp -std=c99 {} -shared -o {}".format(source_code, kernel_file) + compile_command = "gcc -O3 -fPIC -pthread -fopenmp -std=c99 {} -shared -o {}".format( + source_code, kernel_file) print("Compiling", compile_command) exit_state = os.system(compile_command) if not exit_state: @@ -152,7 +158,7 @@ class CPUKernel: print(exception) else: print("Compile default cpu kernel failed, using default cpu kernel code.") - + if kernels is None: # adjust config, use default cpu kernel compile_parallel_kernel = False source_code = default_cpu_kernel_code_path @@ -250,7 +256,8 @@ def compress_int4_weight(weight: torch.Tensor): # (n, m) blockDim, 0, stream, - [ctypes.c_void_p(weight.data_ptr()), ctypes.c_void_p(out.data_ptr()), ctypes.c_int32(n), ctypes.c_int32(m)], + [ctypes.c_void_p(weight.data_ptr()), ctypes.c_void_p(out.data_ptr()), ctypes.c_int32(n), + ctypes.c_int32(m)], ) return out @@ -287,7 +294,8 @@ def extract_weight_to_half(weight: torch.Tensor, scale_list: torch.Tensor, sourc return out -def extract_weight_to_float(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int, quantization_cache=None): +def extract_weight_to_float(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int, + quantization_cache=None): """extract weight on cpu to float32""" if source_bit_width == 8: func = cpu_kernels.int8WeightExtractionFloat @@ -332,7 +340,8 @@ class CacheTensor(): class QuantizedLinear(Linear): - def __init__(self, weight_bit_width: int, weight_tensor=None, bias_tensor=None, quantized_weight=None, quantized_weight_scale=None, quantization_cache=None, empty_init=False, *args, **kwargs): + def __init__(self, weight_bit_width: int, weight_tensor=None, bias_tensor=None, quantized_weight=None, + quantized_weight_scale=None, quantization_cache=None, empty_init=False, *args, **kwargs): super(QuantizedLinear, self).__init__(*args, **kwargs) self.weight_bit_width = weight_bit_width self.quantization_cache = quantization_cache @@ -351,7 +360,8 @@ class QuantizedLinear(Linear): ) self.weight_scale = torch.empty(shape[0], dtype=kwargs["dtype"], device=kwargs["device"]) else: - self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).to(kwargs["dtype"]) + self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).to( + kwargs["dtype"]) self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8) if weight_bit_width == 4: self.weight = compress_int4_weight(self.weight) @@ -370,7 +380,8 @@ class QuantizedLinear(Linear): def forward(self, input): if self.weight.device == torch.device("cpu"): - output = W8A16LinearCPU.apply(input, self.weight, self.weight_scale, self.weight_bit_width, self.quantization_cache) + output = W8A16LinearCPU.apply(input, self.weight, self.weight_scale, self.weight_bit_width, + self.quantization_cache) else: output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width) if self.bias is not None: @@ -386,7 +397,8 @@ class QuantizedLinear(Linear): class QuantizedEmbedding(Embedding): # TODO: backward, check empty_init - def __init__(self, weight_bit_width: int, weight_tensor=None, quantized_weight=None, quantized_weight_scale=None, empty_init=False, *args, **kwargs): + def __init__(self, weight_bit_width: int, weight_tensor=None, quantized_weight=None, quantized_weight_scale=None, + empty_init=False, *args, **kwargs): super(QuantizedEmbedding, self).__init__(*args, **kwargs) self.weight_bit_width = weight_bit_width @@ -404,7 +416,8 @@ class QuantizedEmbedding(Embedding): # TODO: backward, check empty_init ) self.weight_scale = torch.empty(shape[0], dtype=kwargs["dtype"], device=kwargs["device"]) else: - self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).half() + self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).to( + kwargs["dtype"]) self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8) if weight_bit_width == 4: self.weight = compress_int4_weight(self.weight) @@ -414,9 +427,11 @@ class QuantizedEmbedding(Embedding): # TODO: backward, check empty_init def forward(self, input): if self.weight.device == torch.device("cpu"): - original_weight = extract_weight_to_float(weight=self.weight, scale_list=self.weight_scale, source_bit_width=self.weight_bit_width) + original_weight = extract_weight_to_float(weight=self.weight, scale_list=self.weight_scale, + source_bit_width=self.weight_bit_width) else: - original_weight = extract_weight_to_half(weight=self.weight, scale_list=self.weight_scale, source_bit_width=self.weight_bit_width) + original_weight = extract_weight_to_half(weight=self.weight, scale_list=self.weight_scale, + source_bit_width=self.weight_bit_width) output = F.embedding( input, original_weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse @@ -449,7 +464,7 @@ def quantize(model, weight_bit_width, use_quantization_cache=False, empty_init=F current_device = model.device if model.device == torch.device("cpu"): - dtype=torch.float32 + dtype = torch.float32 else: dtype = torch.half