add sensitivity analysis tool for layer-wise FIT and Hessian trace (#592

)

* add FIT and Hessian

* renamed files

torchao/quantization/prototype/mixed_precision/scripts/fit.py

@@ -0,0 +1,105 @@
+import torch
+import numpy as np
+import os
+from tqdm import tqdm
+import transformers
+from datasets import load_dataset
+import random
+from torch.nn.attention import SDPBackend, sdpa_kernel
+
+def get_wikitext2(nsamples, seed, seqlen, tokenizer):
+    traindata = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1", split="train")
+    testdata = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1", split="test")
+
+    trainenc = tokenizer("\n\n".join(traindata["text"]), return_tensors="pt")
+    testenc = tokenizer("\n\n".join(testdata["text"]), return_tensors="pt")
+
+    random.seed(seed)
+    trainloader = []
+    for _ in range(nsamples):
+        i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
+        j = i + seqlen
+        inp = trainenc.input_ids[:, i:j]
+        tar = inp.clone()
+        tar[:, :-1] = -100
+        trainloader.append((inp, tar))
+    return trainloader, testenc
+
+def cal_FIT(device, data, nsamples, model, maxIter, max_seqlen, criterion, num_layers):
+
+    # store the history of trace for each layer
+    estimated_history=[]
+
+    # store the history of mean trace for each layer
+    estimated_mean = [[] for _ in range(num_layers)]
+    trace = [0.] * num_layers
+
+
+    for iteration in range(maxIter):
+        print("iteration: ",iteration)
+        trace_tmp = [0.] * num_layers
+
+        for i in tqdm(range(nsamples)):
+            inputs, targets = data[i]
+            inputs = inputs.to(device)
+            targets = targets.to(device)
+            model.zero_grad()
+            outputs = model(inputs)
+            logits = outputs.logits
+            loss = criterion(logits.view(-1, logits.size(-1)), targets.view(-1))
+
+            grads = torch.autograd.grad(loss, model.parameters())
+
+            # Trace(Fisher Information Matrix) is calculated by the sum of the square of the gradient 
+            for layerid in range(num_layers):
+                for (name, _), grad in zip(model.named_parameters(), grads):
+                    if "."+str(layerid)+"." in name and ("self_attn" in name or "mlp" in name):
+                        trace_tmp[layerid] += torch.sum(grad * grad).item()
+
+            # clean cache
+            model.zero_grad()
+            del grads
+            torch.cuda.empty_cache()
+
+        # calculate the mean of the trace on the calibration dataset
+        for t in range(num_layers):
+            trace[t] = trace_tmp[t] / float(nsamples)
+            estimated_mean[t].append(trace[t])
+
+        print("trace:",trace)
+        estimated_history.append(trace)
+
+    F_average = np.array([np.mean(i) for i in estimated_mean])
+    return F_average, estimated_mean, estimated_history
+
+def main(max_seqlen, checkpoint, nsamples, maxIter, num_layers):
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+    # have been tested models Llama-3-8B, Llama-2-7B, Mistral-7B, and stories110M
+    model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16)
+    tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+    model = model.to(device)
+    model.eval()
+
+    criterion = torch.nn.CrossEntropyLoss()
+
+    # load calibration dataset
+    seed = 0
+    trainloader, testloader = get_wikitext2(nsamples, seed, max_seqlen, tokenizer)
+
+    F_average, estimated_mean, estimated_history = cal_FIT(device=device, data=trainloader, nsamples=nsamples, model=model, maxIter=maxIter, max_seqlen=max_seqlen, criterion=criterion, num_layers=num_layers)
+    print("Iteration Done")
+    print("avg_trace:", F_average)
+    print("estimated_mean:", estimated_mean)
+    print("estimated_history:", estimated_history)
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser(description='Your CLI description.')
+    parser.add_argument('--checkpoint', type=str, default="/home/hanxianhuang/ao/torchao/quantization/prototype/mixed_precision/checkpoints/meta-llama/Meta-Llama-3-8B", help='Path to load model')
+    parser.add_argument('--max_seqlen', type=int, default=2048, help='Max sequence length')
+    parser.add_argument('--maxIter', type=int, default=100, help='The number of iterations to calculate FIT')
+    parser.add_argument('--num_layers', type=int, default=32, help='The number of layers to calculate FIT.')
+    parser.add_argument('--nsamples', type=int, default=128, help='The number of samples in calibration dataset') 
+    args = parser.parse_args()
+    main(args.max_seqlen, args.checkpoint, args.nsamples, args.maxIter, args.num_layers)

torchao/quantization/prototype/mixed_precision/scripts/hessian_grad.py

@@ -0,0 +1,151 @@
+import torch
+import numpy as np
+import os
+from tqdm import tqdm
+import transformers
+from datasets import load_dataset
+import random
+from torch.nn.attention import SDPBackend, sdpa_kernel
+from torch.autograd.functional import hvp
+
+def group_product(xs, ys):
+    return [torch.sum(x * y) for (x, y) in zip(xs, ys)]
+
+def get_wikitext2(nsamples, seed, seqlen, tokenizer):
+    traindata = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1", split="train")
+    testdata = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1", split="test")
+
+    trainenc = tokenizer("\n\n".join(traindata["text"]), return_tensors="pt")
+    testenc = tokenizer("\n\n".join(testdata["text"]), return_tensors="pt")
+
+    random.seed(seed)
+    trainloader = []
+    for _ in range(nsamples):
+        i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
+        j = i + seqlen
+        inp = trainenc.input_ids[:, i:j]
+        tar = inp.clone()
+        tar[:, :-1] = -100
+        trainloader.append((inp, tar))
+    return trainloader, testenc.input_ids
+
+def dataloader_hv_product(layerid, params, device, v, data, nsamples, model, max_seqlen, criterion):
+    model.zero_grad()
+    THv = [torch.zeros(p.size()).to(device) for p in params]  # accumulate result
+
+    # Freeze all the parameters in the model
+    for param in model.parameters():
+        param.requires_grad = False
+
+    # Unfreeze the parameters of attention and MLP layers in layer 0
+    layer_ = model.model.layers[layerid]
+    for param in layer_.self_attn.parameters():
+        param.requires_grad = True
+    for param in layer_.mlp.parameters():
+        param.requires_grad = True
+
+    for i in tqdm(range(nsamples)):
+        torch.cuda.empty_cache()
+        inputs, labels = data[i]
+        inputs = inputs.to(device)
+        labels = labels.to(device)
+        # if use testloader:
+        # inputs = data[:, (i * max_seqlen) : ((i + 1) * max_seqlen)].to(device)
+        # labels = data[:, (i * max_seqlen) : ((i + 1) * max_seqlen)].to(device)
+        model.zero_grad()
+        outputs = model(inputs)
+        logits = outputs.logits
+        loss = criterion(logits.view(-1, logits.size(-1)), labels.view(-1))
+
+        # get the first order gradients
+        grads = torch.autograd.grad(loss, params, create_graph=True, only_inputs=True)
+
+        # calculate Hessian vector product via Jac-vector product
+        Hv = torch.autograd.grad(grads, params, grad_outputs=v, only_inputs=True, retain_graph=False)
+
+        THv = [THv1 + Hv1 + 0.0 for THv1, Hv1 in zip(THv, Hv)]
+
+        # clean cache
+        model.zero_grad()   
+        del Hv
+        del grads                        
+        torch.cuda.empty_cache()
+
+    THv = [THv1 / float(nsamples) for THv1 in THv]
+    return THv
+
+def cal_trace(layerid, params, device, data, nsamples, model, maxIter, max_seqlen, criterion):
+    vhv_c_history = []
+    trace_history = []
+    trace = 0.
+
+    for i in range(maxIter):
+        print("iteration: ",i)
+
+        # generate Rademacher random variables
+        v = [
+            torch.randint_like(p, high=2, device=device)
+            for p in params
+        ]
+
+        for v_i in v:
+            v_i[v_i == 0] = -1
+
+        # calculate Hessian vector product
+        Hv = dataloader_hv_product(layerid, params, device, v, data, nsamples, model, max_seqlen, criterion)
+
+        vHv = group_product(Hv, v)
+
+        vHv_c = np.array([i.cpu().numpy() for i in vHv])
+
+        vhv_c_history.append(vHv_c)
+
+        trace = np.sum(vHv_c)
+
+        trace_history.append(trace)
+        print("trace,", trace)
+        print("trace_history,", trace_history)
+        print("vhv_c_history,", vhv_c_history)
+
+    return np.mean(trace_history)
+
+
+def main(layer_id, checkpoint, max_seqlen, maxIter, nsamples):
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+    # to avoid aten::_scaled_dot_product_flash_attention_backward not implemented error
+    with sdpa_kernel(SDPBackend.MATH):
+
+        # have been tested models Llama-3-8B, Llama-2-7B, Mistral-7B, and stories110M
+        model = transformers.AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16)
+        tokenizer = transformers.AutoTokenizer.from_pretrained(checkpoint)
+        model = model.cuda()
+        model.eval()
+
+        criterion = torch.nn.CrossEntropyLoss()
+
+        # load calibration dataset
+        seed = 0
+        trainloader, testloader = get_wikitext2(128, seed, 2048, tokenizer)
+
+        # calculate Hessian for only one layer each time
+        params=[]
+        layer_ = model.model.layers[layer_id]
+        for param in layer_.self_attn.parameters():
+            params.append(param)
+        for param in layer_.mlp.parameters():
+            params.append(param)
+
+        trace = cal_trace(layerid=layer_id, params=params, device=device, data=trainloader, nsamples=nsamples, model=model, maxIter=maxIter, max_seqlen=max_seqlen, criterion=criterion)
+        print("The trace of layer " + str(layer_id) + " is", trace)
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser(description='Your CLI description.')
+    parser.add_argument('--layer_id', type=int, default=0, help='Which layer to compute the trace and hessian')
+    parser.add_argument('--checkpoint', type=str, default="/home/hanxianhuang/ao/torchao/quantization/prototype/mixed_precision/checkpoints/meta-llama/Meta-Llama-3-8B", help='Path to load model')
+    parser.add_argument('--max_seqlen', type=int, default=2048, help='Max sequence length')
+    parser.add_argument('--maxIter', type=int, default=100, help='The number of iterations to calculate Hessian trace')
+    parser.add_argument('--nsamples', type=int, default=128, help='The number of samples in calibration dataset') 
+    args = parser.parse_args()
+    main(args.layer_id, args.checkpoint, args.max_seqlen, args.maxIter, args.nsamples)