File size: 9,299 Bytes
c32a0aa
 
 
5e16ca3
 
 
c32a0aa
 
 
 
 
 
 
 
 
 
 
 
 
5e16ca3
c32a0aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e16ca3
c32a0aa
 
 
 
 
 
 
 
 
5e16ca3
c32a0aa
 
 
 
 
 
 
 
5e16ca3
 
 
c32a0aa
 
 
 
 
 
5e16ca3
 
 
 
c32a0aa
5e16ca3
 
 
 
 
c32a0aa
5e16ca3
 
 
 
c32a0aa
 
5e16ca3
 
 
 
 
 
 
 
 
 
 
c32a0aa
 
5e16ca3
c32a0aa
 
 
 
 
5e16ca3
c32a0aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e16ca3
 
c32a0aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e16ca3
 
c32a0aa
 
 
 
 
 
 
 
 
5e16ca3
c32a0aa
5e16ca3
c32a0aa
5e16ca3
c32a0aa
 
 
 
 
 
5e16ca3
 
 
 
c32a0aa
5e16ca3
c32a0aa
5e16ca3
c32a0aa
 
 
 
 
 
 
 
5e16ca3
c32a0aa
5e16ca3
c32a0aa
 
 
 
5e16ca3
c32a0aa
5e16ca3
c32a0aa
 
 
5e16ca3
 
 
c32a0aa
5e16ca3
 
 
 
c32a0aa
5e16ca3
c32a0aa
5e16ca3
 
 
 
 
 
 
 
 
 
c32a0aa
 
 
 
 
 
 
5e16ca3
c32a0aa
5e16ca3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
"""
Full benchmark suite comparing:
1. FP16 baseline
2. Uniform 8-bit quantization
3. Naive mixed per-head (uint8 storage β€” not truly packed)
4. Triton mixed per-head (truly packed 4-bit)
Across: memory, speed, perplexity
"""
import torch
import json
import os
import sys
import time
import math
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

sys.path.append(os.path.expanduser("~/kv-hack"))
from kernel.quant_cache import MixedPrecisionKVCache
from kernel.quant_cache_triton import MixedPrecisionKVCacheTriton

# ── config ──────────────────────────────────────────
MODEL_NAME = sys.argv[1] if len(sys.argv) > 1 else "mistral-7b"
MODEL_PATHS = {
    "mistral-7b": "~/kv-hack/mistral-model",
    "llama-3-8b": "~/kv-hack/llama-model",
}
model_path  = os.path.expanduser(MODEL_PATHS[MODEL_NAME])
results_dir = os.path.expanduser(f"~/kv-hack/results/{MODEL_NAME}")

with open(f"{results_dir}/bit_allocation.json") as f:
    bit_alloc_raw = json.load(f)
bit_alloc = {
    int(l): [bit_alloc_raw[l][str(h)]
             for h in range(len(bit_alloc_raw[l]))]
    for l in bit_alloc_raw
}
num_layers = len(bit_alloc)
avg_bits   = sum(b for l in bit_alloc.values() for b in l) / \
             sum(len(l) for l in bit_alloc.values())

print(f"Benchmarking: {MODEL_NAME}")
print(f"Avg bits: {avg_bits:.2f}")
print(f"Theoretical compression: {16/avg_bits:.2f}x")

print("Loading model...")
tokenizer = AutoTokenizer.from_pretrained(model_path)
model     = AutoModelForCausalLM.from_pretrained(
    model_path, dtype=torch.float16, device_map="cuda"
)
model.eval()
print(f"Model loaded: {torch.cuda.memory_allocated()/1e9:.2f} GB")


def measure_kv_compression(context_len: int):
    input_ids = torch.randint(1, 1000, (1, context_len)).cuda()
    with torch.no_grad():
        out = model(input_ids, use_cache=True)
        kv  = out.past_key_values

    fp16_bytes       = 0
    uniform8_bytes   = 0
    naive_real_bytes = 0   # actual GPU bytes for naive (uint8)
    naive_theo_bytes = 0   # theoretical packed size for naive
    triton_bytes     = 0   # actual GPU bytes for triton (truly packed)

    for layer_idx in range(num_layers):
        k = kv.layers[layer_idx].keys
        v = kv.layers[layer_idx].values

        # FP16 baseline
        fp16_bytes     += k.numel() * 2 + v.numel() * 2

        # uniform 8-bit (1 byte per element)
        uniform8_bytes += k.numel() + v.numel()

        # naive mixed precision
        cache_naive = MixedPrecisionKVCache(bit_alloc[layer_idx])
        cache_naive.store(k, v)
        naive_real_bytes += cache_naive.real_gpu_bytes()  # actual GPU
        naive_theo_bytes += cache_naive.memory_bytes()    # theoretical

        # triton true 4-bit
        cache_triton = MixedPrecisionKVCacheTriton(bit_alloc[layer_idx])
        cache_triton.store(k, v)
        triton_bytes += cache_triton.memory_bytes()       # actual GPU (truly packed)

    return {
        "context_len":                  context_len,
        "fp16_mb":                      round(fp16_bytes / 1e6, 2),
        "uniform8_mb":                  round(uniform8_bytes / 1e6, 2),
        "naive_real_gpu_mb":            round(naive_real_bytes / 1e6, 2),
        "naive_theoretical_mb":         round(naive_theo_bytes / 1e6, 2),
        "triton_mb":                    round(triton_bytes / 1e6, 2),
        "naive_real_compression":       round(fp16_bytes / naive_real_bytes, 2),
        "naive_theo_compression":       round(fp16_bytes / naive_theo_bytes, 2),
        "triton_compression_vs_fp16":   round(fp16_bytes / triton_bytes, 2),
        "triton_compression_vs_8bit":   round(uniform8_bytes / triton_bytes, 2),
        "triton_compression_vs_naive":  round(naive_real_bytes / triton_bytes, 2),
    }


def measure_perplexity(num_samples: int = 50):
    print(f"  Computing perplexity on {num_samples} WikiText samples...")
    dataset = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")
    texts   = [t for t in dataset["text"] if len(t.strip()) > 100][:num_samples]

    total_loss   = 0
    total_tokens = 0

    for text in texts:
        inputs = tokenizer(
            text, return_tensors="pt",
            max_length=512, truncation=True
        ).to("cuda")
        if inputs["input_ids"].shape[1] < 10:
            continue
        with torch.no_grad():
            out  = model(**inputs, labels=inputs["input_ids"])
            loss = out.loss.item()
        n = inputs["input_ids"].shape[1]
        total_loss   += loss * n
        total_tokens += n

    return round(math.exp(total_loss / total_tokens), 2)


def measure_speed(context_len: int = 512, n_tokens: int = 100):
    input_ids = torch.randint(1, 1000, (1, context_len)).cuda()
    # warmup
    with torch.no_grad():
        _ = model.generate(
            input_ids, max_new_tokens=10,
            do_sample=False,
            pad_token_id=tokenizer.eos_token_id
        )
    torch.cuda.synchronize()
    t0 = time.time()
    with torch.no_grad():
        _ = model.generate(
            input_ids, max_new_tokens=n_tokens,
            do_sample=False,
            pad_token_id=tokenizer.eos_token_id
        )
    torch.cuda.synchronize()
    return round(n_tokens / (time.time() - t0), 1)


def measure_peak_memory(context_len: int):
    torch.cuda.reset_peak_memory_stats()
    input_ids = torch.randint(1, 1000, (1, context_len)).cuda()
    with torch.no_grad():
        _ = model(input_ids, use_cache=True)
    torch.cuda.synchronize()
    return round(torch.cuda.max_memory_allocated() / 1e9, 2)


# ── RUN ALL BENCHMARKS ───────────────────────────────
print("\n" + "="*75)
print("1. KV CACHE COMPRESSION AT DIFFERENT CONTEXT LENGTHS")
print("="*75)

compression_results = []
for ctx in [512, 1024, 2048, 4096, 8192]:
    print(f"  Context {ctx}...", end=" ", flush=True)
    r = measure_kv_compression(ctx)
    compression_results.append(r)
    print(f"FP16={r['fp16_mb']}MB | "
          f"8bit={r['uniform8_mb']}MB | "
          f"Naive(actual)={r['naive_real_gpu_mb']}MB({r['naive_real_compression']}x) | "
          f"Triton={r['triton_mb']}MB({r['triton_compression_vs_fp16']}x)")

print("\n" + "="*75)
print("2. PEAK GPU MEMORY AT DIFFERENT CONTEXT LENGTHS")
print("="*75)

memory_results = []
for ctx in [1024, 4096, 8192]:
    print(f"  Context {ctx}...", end=" ", flush=True)
    mem = measure_peak_memory(ctx)
    memory_results.append({"context": ctx, "peak_memory_gb": mem})
    print(f"{mem} GB")

print("\n" + "="*75)
print("3. DECODE SPEED")
print("="*75)
print("  Measuring tokens/sec...", end=" ", flush=True)
speed = measure_speed()
print(f"{speed} tokens/sec")

print("\n" + "="*75)
print("4. PERPLEXITY (quality check)")
print("="*75)
perplexity = measure_perplexity(num_samples=50)
print(f"  Perplexity: {perplexity}")

# ── SAVE ─────────────────────────────────────────────
r8k = next(r for r in compression_results if r["context_len"] == 8192)

benchmark_results = {
    "model":                 MODEL_NAME,
    "avg_bits":              round(avg_bits, 2),
    "compression":           compression_results,
    "memory":                memory_results,
    "decode_tokens_per_sec": speed,
    "perplexity":            perplexity,
    "summary": {
        "fp16_8k_mb":                  r8k["fp16_mb"],
        "uniform8_8k_mb":              r8k["uniform8_mb"],
        "naive_real_8k_mb":            r8k["naive_real_gpu_mb"],
        "naive_theoretical_8k_mb":     r8k["naive_theoretical_mb"],
        "triton_8k_mb":                r8k["triton_mb"],
        "naive_real_compression_8k":   r8k["naive_real_compression"],
        "naive_theo_compression_8k":   r8k["naive_theo_compression"],
        "triton_compression_8k":       r8k["triton_compression_vs_fp16"],
        "triton_vs_naive_8k":          r8k["triton_compression_vs_naive"],
        "triton_vs_8bit_8k":           r8k["triton_compression_vs_8bit"],
    }
}

out_path = f"{results_dir}/benchmark_results.json"
with open(out_path, "w") as f:
    json.dump(benchmark_results, f, indent=2)

print("\n" + "="*75)
print("SUMMARY")
print("="*75)
print(f"Model:                    {MODEL_NAME}")
print(f"Avg bits per head:        {avg_bits:.2f}")
print(f"Perplexity:               {perplexity}")
print(f"Decode speed:             {speed} tokens/sec")
print()
print(f"KV Cache at 8K context:")
print(f"  FP16 baseline:          {r8k['fp16_mb']} MB       (1.00x)")
print(f"  Uniform 8-bit:          {r8k['uniform8_mb']} MB     (2.00x)")
print(f"  Naive per-head (actual GPU): {r8k['naive_real_gpu_mb']} MB   ({r8k['naive_real_compression']}x)  ← uint8 storage")
print(f"  Naive per-head (theoretical): {r8k['naive_theoretical_mb']} MB  ({r8k['naive_theo_compression']}x) ← if truly packed")
print(f"  Triton true 4-bit:      {r8k['triton_mb']} MB   ({r8k['triton_compression_vs_fp16']}x)  ← actual GPU")
print(f"  Triton vs Naive:        {r8k['triton_compression_vs_naive']}x smaller on GPU")
print(f"  Triton vs 8-bit:        {r8k['triton_compression_vs_8bit']}x smaller")
print(f"\nβœ… Saved to {out_path}")