Merge pull request #416 from turboderp/dev

Merge dev branch

Author: turboderp

conversion/adaptivegptq.py

@@ -69,10 +69,14 @@ def find_params(self, x):
         self.scale = qscale_tw * best_p
         self.qscale_max = qscale_max_t * best_p
 
+        # Make sure scales are rounded correctly for sanity test
+        prescale = torch.tensor([1 / 256], dtype = torch.half, device = self.scale.device)
+        self.scale = ((self.qscale * self.qscale).to(torch.half) * (self.qscale_max.half() * prescale)).float()
+
 
 class AdaptiveGPTQ:
 
-    percdamp: float = 0.07
+    percdamp: float = 0.12
 
     layer: nn.Linear
     device: torch.device

conversion/measure.py

@@ -387,7 +387,7 @@ def measure_quant(job, save_fn, model):
     overall_rolling_accuracy = 0  
 
     last_snapshot_time = time.time()
-    snapshot_interval_s = 90
+    snapshot_interval_s = 180
 
     temp_filename = os.path.join(job["out_dir"], "hidden_states_temp.safetensors")
     states_filename = os.path.join(job["out_dir"], "hidden_states.safetensors")

conversion/optimize.py

@@ -1,6 +1,8 @@
 from conversion.qparams import QParams
+from exllamav2.ext import exllamav2_ext as ext_c, none_tensor
 import math
 import itertools
+import time
 
 def optimize(job, save_fn, model):
 
@@ -9,11 +11,19 @@ def optimize(job, save_fn, model):
     mlp_key_up = model.config.arch.mlp_key_up
     mlp_key_down = model.config.arch.mlp_key_down
 
-    error_norm = 2.4
-    max_step_size = 2
-    first_layer_bias = 10
-    bias_layers = 2
-    bias_iter = 10
+    norm_interval = (1.5, 3.5)
+    norm_2ndstage = 0.15
+    anneal_temp_max = 2
+    anneal_temp_min = 0.0001
+    anneal_cooling_factor = 0.995
+    anneal_iter = 1000
+    anneal_samples = 80
+    anneal_stages = 3
+
+    # max_step_size = 2
+    # first_layer_bias = 4
+    # bias_layers = 2
+    # bias_iter = 0
 
     key = "model.layers.0"
     key_q = key + ".self_attn.q_proj"
@@ -57,21 +67,14 @@ def optimize(job, save_fn, model):
     numel = sum(m.numel() for m in model.modules[1 : num_modules + 1])
 
     target_bpw = job["bits"]
-    weight_budget = numel * target_bpw
+    weight_budget = int(numel * target_bpw)
 
     # Compile options
 
     measurement = job["measurement"]
-
-    def fn(x, idx):
-        if idx < bias_layers:
-            return 1 - ((1 - x) ** error_norm) * first_layer_bias
-        else:
-            return 1 - ((1 - x) ** error_norm)
-
-    weights = []
-    values = []
+    slots = []
     params = []
+
     for i in range(num_layers):
         if model.config.arch.parallel_decoder_blocks:
             m1 = measurement["model.layers." + str(i) + ".parallel_decoder"]["attn"]
@@ -80,162 +83,83 @@ def fn(x, idx):
             m1 = measurement["model.layers." + str(i) + ".self_attn"]
             m2 = measurement["model.layers." + str(i) + "." + mlp_mode]
         for m in [m1, m2]:
-            v = [fn(e["accuracy"], i) for e in m]
-            w = [e["total_bits"] for e in m]
-            weights.append(w)
-            values.append(v)
-            params.append(m)
-
-    print(" -- Pruning...")
-
-    # Sort options by weight, eliminate strictly worse options
-
-    for i in range(num_layers * 2):
-        combined = sorted(zip(weights[i], values[i], params[i]))
-        w_, v_, p_ = zip(*combined)
-        w_ = list(w_)
-        v_ = list(v_)
-        p_ = list(p_)
-        j = 1
-        while j < len(v_):
-            if v_[j] <= v_[j - 1]:
-                w_.pop(j)
-                v_.pop(j)
-                p_.pop(j)
-            else:
-                j += 1
-        weights[i] = w_
-        values[i] = v_
-        params[i] = p_
-
-    # Quick and dirty iterative solver
-
-    print(" -- Solving...")
-
-    f_solution = [0] * num_layers * 2
-    weight = sum(weights[i][0] for i in range(num_layers * 2))
-    value = 1
-    for i in range(num_layers * 2): value *= values[i][0]
-
-    iteration = 0
-
-    while True:
-        min_idx = -1
-        min_value = float("inf")
-        iteration += 1
-        for i in range(bias_layers if iteration < bias_iter else num_layers * 2):
-            s = f_solution[i]
-            if values[i][s] < min_value:
-                if s < len(weights[i]) - 1:
-                    added_w = weights[i][s + 1] - weights[i][s]
-                    if added_w + weight <= weight_budget:
-                        min_idx = i
-                        min_value = values[i][s]
-        if min_idx == -1: break
-        s = f_solution[min_idx]
-        weight += weights[min_idx][s + 1] - weights[min_idx][s]
-        value *= values[min_idx][s + 1] / values[min_idx][s]
-        f_solution[min_idx] += 1
-
-    bpw = weight / numel
-    print(f" -- Score: {value:.8f}  bpw: {bpw:.4f}")
-
-    def improve(solution, s_weight, hold = None):
-
-        if hold is None: hold = []
-        best_idx = -1
-        best_ratio = 0
-        best_add_w = 0
-        best_add_v = 0
-        for idx in range(num_layers * 2):
-            if idx in hold: continue
-
-            si = solution[idx]
-            if si == len(weights[idx]) - 1: continue
-
-            add_w = weights[idx][si + 1] - weights[idx][si]
-            if s_weight + add_w > weight_budget: continue
-
-            add_v = values[idx][si + 1] / values[idx][si]
-            ratio = add_v / add_w
-            if ratio > best_ratio:
-                best_ratio = ratio
-                best_idx = idx
-                best_add_w = add_w
-                best_add_v = add_v
-
-        return best_idx, best_add_w, best_add_v
-
-    # while True:
-    #     b_idx, b_add_w, b_add_v = improve(f_solution, weight)
-    #     if b_idx == -1:
-    #         break
-    #
-    #     f_solution[b_idx] += 1
-    #     weight += b_add_w
-    #     value += b_add_v
-    #
-    # bpw = weight / numel
-    # print(f" -- Score: {math.exp(value):.8f}  bpw: {bpw:.4f}")
-
-    best_value = value
-    prev_best_value = value
-    step_size = 1
-
-    while True:
-
-        for i, j in itertools.permutations(range(num_layers * 2), 2):
-
-            t_solution = f_solution.copy()
-            t_solution[i] = max(t_solution[i] - step_size, 0)
-            t_solution[j] = max(t_solution[j] - step_size, 0)
-
-            t_weight = sum(weights[k][t_solution[k]] for k in range(num_layers * 2))
-            t_value = 1
-            for k in range(num_layers * 2): t_value *= values[k][t_solution[k]]
-
-            while True:
-                b_idx, b_add_w, b_add_v = improve(t_solution, t_weight, [i, j])
-                if b_idx == -1:
-                    break
-                t_solution[b_idx] += 1
-                t_weight += b_add_w
-                t_value *= b_add_v
-
-            if t_value > best_value:
-                f_solution = t_solution
-                best_value = t_value
-                break
-
-        if best_value == prev_best_value:
-            step_size += 1
-            if step_size > max_step_size: break
-            continue
-
-        bpw = t_weight / numel
-        print(f" -- Score: {best_value:.8f}  bpw: {bpw:.4f}")
-        prev_best_value = best_value
+            slot = []
+            param = []
+            for opt in m:
+                o = (int(opt["total_bits"]), 1 - opt["accuracy"])
+                slot.append(o)
+                param.append(opt)
+            slots.append(slot)
+            params.append(param)
+
+    # Find some solutions
+
+    last_update = 0
+    m = float("inf")
+    p = float("inf")
+    for i in range(anneal_stages * anneal_samples):
+        if time.time() - last_update > 1 or i == anneal_samples - 1:
+            print(f" -- Optimizing: {i + 1:4}/{anneal_stages * anneal_samples:4}")
+            last_update = time.time()
+
+        if i < anneal_samples:
+            t = i / (anneal_samples - 1)
+            norm = (1 - t) * norm_interval[0] + t * norm_interval[1]
+
+        elif i < anneal_samples * 2:
+            if i == anneal_samples:
+                norm_a = bestnorm - norm_2ndstage / 2
+                norm_b = bestnorm + norm_2ndstage / 2
+            t = i / (anneal_samples - 1) - 1
+            norm = (1 - t) * norm_a + t * norm_b
+
+        else:
+            norm = bestnorm
+
+        s_, si_, p_, c_, m_ = ext_c.sim_anneal(slots,
+                                               weight_budget,
+                                               anneal_temp_max,
+                                               anneal_cooling_factor,
+                                               anneal_temp_min,
+                                               anneal_iter,
+                                               norm)
+
+        if i < anneal_samples * 2:
+            if m_ < m:
+                m = m_
+                bestnorm = norm
+        else:
+            if p_ < p:
+                s, si, p, m = s_, si_, p_, m_
+
+    solution_idx = si
+    print(f" -- max(err): {m:.6f}")
+    print(f" -- error_norm: {bestnorm:.6f}")
+
 
     # Save strategy
 
     print(" -- Quantization strategy:")
 
-    errp = 1
+    logerr = 0
+    maxerr = 0
     job["strategy"] = {}
     for layer_ in range(num_layers):
 
         k1 = "model.layers." + str(layer_) + ".self_attn"
         k2 = "model.layers." + str(layer_) + "." + mlp_mode
-        p1 = params[layer_ * 2][f_solution[layer_ * 2]]
-        p2 = params[layer_ * 2 + 1][f_solution[layer_ * 2 + 1]]
+        p1 = params[layer_ * 2][solution_idx[layer_ * 2]]
+        p2 = params[layer_ * 2 + 1][solution_idx[layer_ * 2 + 1]]
 
         for (k, p, n) in zip((k1, k2), (p1, p2), (numel_attn, numel_mlp)):
             job["strategy"][k] = p
             bpw = p["total_bits"] / n
             err = 1 - p["accuracy"]
             print(f" --   {k:50} {bpw:1.4f} bpw - exp. error: {err:1.8f}")
-            errp *= (1 - err)
+            logerr += math.log(err)
+            maxerr = max(err, maxerr)
 
-    print(f" -- Total exp. error: {1 - errp:1.12f}")
+    print(f" -- sum(log(err)): {logerr:.6f}")
+    print(f" -- max(err): {maxerr:.6f}")
 
     xx = 0

conversion/quantize.py

@@ -253,7 +253,7 @@ def quant_parallel_decoder(job, module, hidden_states, target_states, quantizers
 def quant(job, save_fn, model):
 
     last_snapshot_time = time.time()
-    snapshot_interval_s = 90
+    snapshot_interval_s = 180
 
     temp_filename = os.path.join(job["out_dir"], "hidden_states_temp.safetensors")
     states_filename = os.path.join(job["out_dir"], "hidden_states.safetensors")
@@ -526,4 +526,4 @@ def quant(job, save_fn, model):
             del job["invalid"]
             save_fn()
 
-            time_since_snapshot = time.time()
+            last_snapshot_time = time.time()

convert.py

@@ -8,6 +8,7 @@
 from conversion.optimize import optimize
 from conversion.compile import compile_model
 from conversion.qparams import qparams_headoptions
+import torch
 
 parser = argparse.ArgumentParser(description = "Convert model to ExLlamaV2")
 parser.add_argument("-i", "--in_dir", type = str, help = "Input directory", default = "")
@@ -29,6 +30,8 @@
 
 args = parser.parse_args()
 
+torch.set_printoptions(precision = 7, sci_mode = False, linewidth = 200)
+
 # Check some args
 
 if not args.in_dir:

doc/qcache_eval.md

@@ -15,20 +15,23 @@ The tl;dr:
 Token-level perplexity tests for various full-precision and quantized models using FP16, FP8 and Q4 cache
 modes. Dataset is The Pile, 10 rows of 512 tokens per test. 
 
-Model	| Precision	| FP16 cache	| FP8 cache	| Q4 cache
---------|-----------|---------------|-----------|---------
-Mistral 7B Instruct	| 3.0 bpw	| 13.33	| 13.43	| 13.41
---	| 3.5 bpw	| 13.07	| 13.14	| 13.12
---	| 4.0 bpw	| 12.90	| 12.90	| 12.90
---	| 5.0 bpw	| 12.73	| 12.73	| 12.75
---	| 6.0 bpw	| 12.73	| 12.75	| 12.74
---	| FP16	| 12.69	| 12.71	| 12.72
-Mixtral 8x7B	| 3.5 bpw	| 10.27	| 10.41	| 10.39
---	| 4.0 bpw	| 10.09	| 10.26	| 10.23
---	| 5.0 bpw	| 10.02	| 10.16	| 10.15
-Llama2 7B	| 4.0 bpw	| 11.43	| 11.92	| 11.74
---	| 5.0 bpw	| 11.13	| 11.40	| 11.31
---	| FP16	| 10.91	| 11.24	| 11.16
+Results are updated for the new method which uses Hadamard rotations on the keys/values. Old results for version
+0.0.18 and prior kept for reference.
+
+Model	| Precision	 | FP16 cache	 | FP8 cache	| Q4 cache (old) | Q4 cache
+--------|---------|-------------|-----------|-------|----------
+Mistral 7B Instruct	| 3.0 bpw | **13.33**	  | 13.43	| 13.41 | **13.37**
+--	| 3.5 bpw	 | **13.07**	  | 13.14	| 13.12 | **13.09**
+--	| 4.0 bpw	 | **12.90**	  | 12.90	| 12.90 | **12.90**
+--	| 5.0 bpw	 | **12.73**	  | 12.73	| 12.75 | **12.75** 
+--	| 6.0 bpw	 | **12.73**	  | 12.75	| 12.74 | **12.74**
+--	| FP16	   | **12.69**	  | 12.71	| 12.72 | **12.69**
+Mixtral 8x7B	| 3.5 bpw	 | **10.27**	  | 10.41	| 10.39 | **10.32** 
+--	| 4.0 bpw	 | **10.09**	  | 10.26	| 10.23 | **10.19**
+--	| 5.0 bpw	 | **10.02**	  | 10.16	| 10.15 | **10.04**
+Llama2 7B	| 4.0 bpw	 | **11.43**	  | 11.92	| 11.74 | **11.60** 
+--	| 5.0 bpw	 | **11.13**	  | 11.40	| 11.31 | **11.19**
+--	| FP16	   | **10.91**	  | 11.24	| 11.16 | **11.05**
 
 
 ### HumanEval
@@ -37,6 +40,8 @@ The following are HumanEval tests on various full-precision and quantized models
 respectively. Number of samples per task is limited to 10 (still giving 39360 completions in total produced
 over about 24 hours.)
 
+The following tests were done prior to the improvements in 0.0.18-dev. 
+
 #### pass@1 
 
 Model |	Precision	| FP16 cache  |	Q4 cache	| diff