Text Generation
Transformers
Safetensors
phi3
quantized
quantport
smoothquant
wikitext-2
conversational
text-generation-inference
8-bit precision
Instructions to use chris320211/smoothquant-phi35-mini-vastrtx3090 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use chris320211/smoothquant-phi35-mini-vastrtx3090 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="chris320211/smoothquant-phi35-mini-vastrtx3090") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("chris320211/smoothquant-phi35-mini-vastrtx3090") model = AutoModelForCausalLM.from_pretrained("chris320211/smoothquant-phi35-mini-vastrtx3090", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use chris320211/smoothquant-phi35-mini-vastrtx3090 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "chris320211/smoothquant-phi35-mini-vastrtx3090" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "chris320211/smoothquant-phi35-mini-vastrtx3090", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/chris320211/smoothquant-phi35-mini-vastrtx3090
- SGLang
How to use chris320211/smoothquant-phi35-mini-vastrtx3090 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "chris320211/smoothquant-phi35-mini-vastrtx3090" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "chris320211/smoothquant-phi35-mini-vastrtx3090", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "chris320211/smoothquant-phi35-mini-vastrtx3090" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "chris320211/smoothquant-phi35-mini-vastrtx3090", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use chris320211/smoothquant-phi35-mini-vastrtx3090 with Docker Model Runner:
docker model run hf.co/chris320211/smoothquant-phi35-mini-vastrtx3090
File size: 32,754 Bytes
99a8c10 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 | diff --git a/quant_agent_inference_adapter.py b/quant_agent_inference_adapter.py
new file mode 100644
index 0000000..392f991
--- /dev/null
+++ b/quant_agent_inference_adapter.py
@@ -0,0 +1,456 @@
+"""Reload a SmoothQuant W8A8 int8 artifact (``w8a8_int8`` format) and run it on int8.
+
+The artifact (written by smoothquant/int8_export.py) stores, for every quantized
+decoder linear, an int8 ``weight`` [out, in] and an fp16 per-output-channel
+``weight_scale`` [out, 1], plus fp16 smoothed norms, embeddings, final norm and
+(untied) lm_head. Smoothing is folded into those weights offline, so no
+transform runs at load or inference time and there are no eval-only flags.
+
+The kernels below are copied verbatim from the shared library entry
+``agents/skills/_shared/kernels/w8a8_int8/triton_w8a8.py`` (``w8a8_int8_triton``,
+matched by this artifact's quantization_config.json): a Triton per-token absmax
+int8 activation quant, then a Triton int8 x int8 -> int32 tensor-core GEMM with a
+fused ``acc * act_scale[m] * weight_scale[n] (+ bias[n])`` epilogue. On CPU, or
+with ``QUANT_AGENT_W8A8_REFERENCE=1``, each linear falls back to the dense fp16
+dequant reference.
+
+The model is always built with the native transformers class and SDPA
+attention: the artifact has no ``auto_map`` and no modeling code, so remote code
+is never loaded whatever ``trust_remote_code`` says.
+"""
+import glob
+import json
+import os
+import time
+
+import torch
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+from torch import nn
+
+
+# --- kernels: verbatim copy of agents/skills/_shared/kernels/w8a8_int8/triton_w8a8.py ---
+
+@triton.jit
+def _act_quant_kernel(
+ x_ptr, q_ptr, s_ptr, K, stride_xm, stride_qm,
+ BLOCK_K: tl.constexpr,
+):
+ """Per-token absmax int8 quantization of one row of x (fp16/bf16)."""
+ row = tl.program_id(0).to(tl.int64)
+ offs = tl.arange(0, BLOCK_K)
+ mask = offs < K
+ x_raw = tl.load(x_ptr + row * stride_xm + offs, mask=mask, other=0.0)
+ act_ty = x_raw.dtype
+ x = x_raw.to(tl.float32)
+ amax = tl.max(tl.abs(x), axis=0)
+ # Mirror torch on the activation dtype: clamp, cast, divide, cast.
+ amax = tl.maximum(amax, 1e-5).to(act_ty).to(tl.float32)
+ scale = tl.math.div_rn(amax, 127.0).to(act_ty).to(tl.float32)
+ v = tl.math.div_rn(x, scale).to(act_ty).to(tl.float32)
+ # Round half to even, like torch.round; |v| <= ~127.5 so v + 0.5 is exact.
+ r = tl.floor(v + 0.5)
+ tie = (r - v) == 0.5
+ odd = (r - 2.0 * tl.floor(r * 0.5)) != 0.0
+ r = tl.where(tie & odd, r - 1.0, r)
+ r = tl.minimum(tl.maximum(r, -127.0), 127.0)
+ tl.store(q_ptr + row * stride_qm + offs, r.to(tl.int8), mask=mask)
+ tl.store(s_ptr + row, scale)
+
+
+@triton.jit
+def _act_quant_looped_kernel(
+ x_ptr, q_ptr, s_ptr, K, stride_xm, stride_qm,
+ BLOCK_K: tl.constexpr,
+):
+ """Same math as _act_quant_kernel for long rows: absmax pass, then quantize pass."""
+ row = tl.program_id(0).to(tl.int64)
+ x_row = x_ptr + row * stride_xm
+ q_row = q_ptr + row * stride_qm
+ offs = tl.arange(0, BLOCK_K)
+ run_max = tl.zeros((BLOCK_K,), dtype=tl.float32)
+ for k0 in range(0, K, BLOCK_K):
+ mask = (k0 + offs) < K
+ x = tl.load(x_row + k0 + offs, mask=mask, other=0.0).to(tl.float32)
+ run_max = tl.maximum(run_max, tl.abs(x))
+ act_ty = x_ptr.dtype.element_ty
+ amax = tl.max(run_max, axis=0)
+ amax = tl.maximum(amax, 1e-5).to(act_ty).to(tl.float32)
+ scale = tl.math.div_rn(amax, 127.0).to(act_ty).to(tl.float32)
+ for k0 in range(0, K, BLOCK_K):
+ mask = (k0 + offs) < K
+ x = tl.load(x_row + k0 + offs, mask=mask, other=0.0).to(tl.float32)
+ v = tl.math.div_rn(x, scale).to(act_ty).to(tl.float32)
+ r = tl.floor(v + 0.5)
+ tie = (r - v) == 0.5
+ odd = (r - 2.0 * tl.floor(r * 0.5)) != 0.0
+ r = tl.where(tie & odd, r - 1.0, r)
+ r = tl.minimum(tl.maximum(r, -127.0), 127.0)
+ tl.store(q_row + k0 + offs, r.to(tl.int8), mask=mask)
+ tl.store(s_ptr + row, scale)
+
+
+@triton.jit
+def _w8a8_gemm_kernel(
+ a_ptr, b_ptr, c_ptr, as_ptr, bs_ptr, bias_ptr,
+ M, N, K,
+ stride_am, stride_ak, stride_bn, stride_bk, stride_cm, stride_cn,
+ HAS_BIAS: tl.constexpr,
+ EVEN_K: tl.constexpr,
+ BLOCK_M: tl.constexpr,
+ BLOCK_N: tl.constexpr,
+ BLOCK_K: tl.constexpr,
+ GROUP_M: tl.constexpr,
+):
+ """C[M, N] = (A_int8[M, K] @ B_int8[N, K]^T) * a_scale[M] * b_scale[N] + bias[N]."""
+ pid = tl.program_id(0)
+ num_pid_m = tl.cdiv(M, BLOCK_M)
+ num_pid_n = tl.cdiv(N, BLOCK_N)
+ num_pid_in_group = GROUP_M * num_pid_n
+ group_id = pid // num_pid_in_group
+ first_pid_m = group_id * GROUP_M
+ group_size_m = min(num_pid_m - first_pid_m, GROUP_M)
+ pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
+ pid_n = (pid % num_pid_in_group) // group_size_m
+
+ offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+ offs_n = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
+ # Wrap out-of-range rows/cols onto valid memory; the store is masked.
+ offs_am = offs_m % M
+ offs_bn = offs_n % N
+ offs_k = tl.arange(0, BLOCK_K)
+ a_ptrs = a_ptr + offs_am[:, None].to(tl.int64) * stride_am + offs_k[None, :] * stride_ak
+ b_ptrs = b_ptr + offs_bn[None, :].to(tl.int64) * stride_bn + offs_k[:, None] * stride_bk
+
+ acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.int32)
+ for k in range(0, tl.cdiv(K, BLOCK_K)):
+ if EVEN_K:
+ a = tl.load(a_ptrs)
+ b = tl.load(b_ptrs)
+ else:
+ k_left = K - k * BLOCK_K
+ a = tl.load(a_ptrs, mask=offs_k[None, :] < k_left, other=0)
+ b = tl.load(b_ptrs, mask=offs_k[:, None] < k_left, other=0)
+ acc = tl.dot(a, b, acc, out_dtype=tl.int32)
+ a_ptrs += BLOCK_K * stride_ak
+ b_ptrs += BLOCK_K * stride_bk
+
+ a_scale = tl.load(as_ptr + offs_am).to(tl.float32)
+ b_scale = tl.load(bs_ptr + offs_bn).to(tl.float32)
+ out = acc.to(tl.float32) * a_scale[:, None] * b_scale[None, :]
+ if HAS_BIAS:
+ bias = tl.load(bias_ptr + offs_bn).to(tl.float32)
+ out = out + bias[None, :]
+ c_ptrs = c_ptr + offs_m[:, None].to(tl.int64) * stride_cm + offs_n[None, :] * stride_cn
+ c_mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
+ tl.store(c_ptrs, out.to(c_ptr.dtype.element_ty), mask=c_mask)
+
+
+def quantize_activation_int8(x):
+ """x [M, K] fp16/bf16 contiguous -> (int8 codes [M, K], fp32 per-token scale [M])."""
+ m, k = x.shape
+ q = torch.empty((m, k), dtype=torch.int8, device=x.device)
+ scale = torch.empty((m,), dtype=torch.float32, device=x.device)
+ if m == 0:
+ return q, scale
+ block_k = triton.next_power_of_2(int(k))
+ args = (x, q, scale, int(k), int(x.stride(0)), int(q.stride(0)))
+ if block_k <= 4096:
+ # Whole row in registers: one read of x.
+ _act_quant_kernel[(int(m),)](*args, BLOCK_K=block_k, num_warps=8)
+ else:
+ # Long rows: absmax pass + quantize pass over 4096-wide chunks; the
+ # second read hits L2. Measured at copy bandwidth on sm_86.
+ _act_quant_looped_kernel[(int(m),)](*args, BLOCK_K=4096, num_warps=8)
+ return q, scale
+
+
+def _gemm_config(m, n, k):
+ """Default (BLOCK_M, BLOCK_N, BLOCK_K, GROUP_M, num_warps, num_stages).
+
+ Tuned on sm_86 (RTX 3090, see kernel.json `tuned_on`). On another compute
+ capability, run `tune_config` on the model's shapes and pass the result as
+ `config` to `w8a8_matmul`.
+ """
+ if m <= 16:
+ return 16, 64, 256, 1, 4, 3
+ if m <= 64:
+ return 64, 64, 128, 8, 4, 4
+ return 128, 128, 64, 8, 4, 4
+
+
+def w8a8_matmul(q, a_scale, weight, weight_scale, bias, out_dtype, config=None):
+ """int8 q [M, K] x int8 weight [N, K] -> out_dtype [M, N] with the dequant epilogue."""
+ m, k = q.shape
+ n = weight.shape[0]
+ out = torch.empty((m, n), dtype=out_dtype, device=q.device)
+ if m == 0:
+ return out
+ block_m, block_n, block_k, group_m, warps, stages = config or _gemm_config(m, n, k)
+ grid = (triton.cdiv(int(m), block_m) * triton.cdiv(int(n), block_n),)
+ _w8a8_gemm_kernel[grid](
+ q, weight, out, a_scale, weight_scale, bias if bias is not None else a_scale,
+ int(m), int(n), int(k),
+ int(q.stride(0)), int(q.stride(1)),
+ int(weight.stride(0)), int(weight.stride(1)),
+ int(out.stride(0)), int(out.stride(1)),
+ HAS_BIAS=bias is not None,
+ EVEN_K=(int(k) % block_k == 0),
+ BLOCK_M=block_m, BLOCK_N=block_n, BLOCK_K=block_k, GROUP_M=group_m,
+ num_warps=warps, num_stages=stages,
+ )
+ return out
+
+
+def w8a8_linear(x2d, weight, weight_scale, bias):
+ """Real int8 W8A8 linear: Triton per-token quant, then int8 GEMM + epilogue."""
+ q, a_scale = quantize_activation_int8(x2d)
+ return w8a8_matmul(q, a_scale, weight, weight_scale, bias, x2d.dtype)
+
+
+def w8a8_linear_reference(x, weight, weight_scale, bias):
+ """fp16 dequant reference: same quantization, dense fp16 matmul. CPU-safe."""
+ x_scale = x.abs().amax(dim=-1, keepdim=True).clamp(min=1e-5) / 127
+ x_q = torch.round(x / x_scale).clamp_(-127, 127).mul_(x_scale)
+ w = weight.to(x.dtype).mul_(weight_scale.reshape(-1, 1).to(x.dtype))
+ b = None if bias is None else bias.to(x.dtype)
+ return F.linear(x_q, w, b)
+
+
+CONFIG_CANDIDATES = (
+ (128, 128, 64, 8, 4, 4),
+ (128, 128, 64, 8, 8, 3),
+ (128, 256, 64, 8, 8, 3),
+ (64, 128, 128, 8, 4, 4),
+ (64, 64, 128, 8, 4, 4),
+ (16, 64, 256, 1, 4, 3),
+)
+
+
+def tune_config(m, n, k, *, device="cuda", candidates=CONFIG_CANDIDATES, iters=20):
+ """Time each candidate tile config on random int8 inputs of this shape; return the fastest.
+
+ Use once per (M, N, K) on a GPU whose compute capability is not in kernel.json
+ `tuned_on`, then pass the result as `config` to `w8a8_matmul`.
+ """
+ q = torch.randint(-127, 128, (m, k), dtype=torch.int8, device=device)
+ w = torch.randint(-127, 128, (n, k), dtype=torch.int8, device=device)
+ a_scale = torch.rand(m, dtype=torch.float32, device=device)
+ w_scale = torch.rand(n, 1, dtype=torch.float16, device=device)
+ best, best_t = None, float("inf")
+ for config in candidates:
+ try:
+ w8a8_matmul(q, a_scale, w, w_scale, None, torch.float16, config)
+ except Exception: # a tile the GPU cannot fit (shared memory, registers)
+ continue
+ torch.cuda.synchronize(device)
+ start = time.perf_counter()
+ for _ in range(iters):
+ w8a8_matmul(q, a_scale, w, w_scale, None, torch.float16, config)
+ torch.cuda.synchronize(device)
+ elapsed = (time.perf_counter() - start) / iters
+ if elapsed < best_t:
+ best, best_t = config, elapsed
+ if best is None:
+ raise RuntimeError(f"no candidate config ran for M={m} N={n} K={k}")
+ return best
+
+
+# --- loader (port-specific) ----------------------------------------------------
+
+QUANT_AGENT_ADAPTER_API = 1
+QUANT_CONFIG_NAME = "quantization_config.json"
+EXPECTED_FORMAT = "w8a8_int8"
+SCALE_SUFFIX = ".weight_scale"
+REFERENCE_ENV = "QUANT_AGENT_W8A8_REFERENCE"
+
+
+def _w8a8_int8_linear_class():
+ class W8A8Int8Linear(nn.Module):
+ """int8 weight + per-channel scale; per-token dynamic int8 activations; int8 GEMM."""
+
+ def __init__(self, in_features, out_features, has_bias, dtype):
+ super().__init__()
+ self.in_features = in_features
+ self.out_features = out_features
+ self.register_buffer(
+ "weight", torch.empty(out_features, in_features, dtype=torch.int8)
+ )
+ self.register_buffer(
+ "weight_scale", torch.empty(out_features, 1, dtype=dtype)
+ )
+ if has_bias:
+ self.register_buffer("bias", torch.empty(out_features, dtype=dtype))
+ else:
+ self.bias = None
+ self.use_reference = os.environ.get(REFERENCE_ENV) == "1"
+
+ @torch.no_grad()
+ def forward(self, x):
+ if (
+ self.use_reference
+ or not x.is_cuda
+ or x.dtype not in (torch.float16, torch.bfloat16)
+ ):
+ return w8a8_linear_reference(x, self.weight, self.weight_scale, self.bias)
+ lead = x.shape[:-1]
+ x2d = x.reshape(-1, self.in_features)
+ if not x2d.is_contiguous():
+ x2d = x2d.contiguous()
+ out = w8a8_linear(x2d, self.weight, self.weight_scale, self.bias)
+ return out.reshape(*lead, self.out_features)
+
+ def extra_repr(self):
+ return (
+ f"in_features={self.in_features}, out_features={self.out_features}, "
+ f"bias={self.bias is not None}, weight=int8 per_channel, "
+ "act=int8 per_token dynamic, kernel=triton int8 gemm"
+ )
+
+ return W8A8Int8Linear
+
+
+def _resolve_dtype(dtype):
+ if dtype in (torch.bfloat16, "bfloat16", "bf16"):
+ return torch.bfloat16
+ if dtype in (None, "auto", torch.float16, "float16", "fp16", "half"):
+ return torch.float16
+ raise ValueError(f"unsupported dtype for the W8A8 int8 artifact: {dtype!r}")
+
+
+def _read_quant_config(model_path):
+ path = os.path.join(model_path, QUANT_CONFIG_NAME)
+ if not os.path.isfile(path):
+ raise FileNotFoundError(f"{QUANT_CONFIG_NAME} is missing in {model_path}")
+ with open(path) as handle:
+ qcfg = json.load(handle)
+ if qcfg.get("format") != EXPECTED_FORMAT:
+ raise ValueError(f"expected format {EXPECTED_FORMAT!r}, got {qcfg.get('format')!r}")
+ if str(qcfg.get("act_quant")) != "per_token" or str(qcfg.get("weight_quant")) != "per_channel":
+ raise ValueError(
+ "this adapter handles per_channel int8 weights with per_token dynamic activations only"
+ )
+ return qcfg
+
+
+def _read_state(model_path):
+ from safetensors.torch import load_file
+
+ files = sorted(glob.glob(os.path.join(model_path, "*.safetensors")))
+ if not files:
+ raise FileNotFoundError(f"no .safetensors weights in {model_path}")
+ state = {}
+ for path in files:
+ part = load_file(path, device="cpu")
+ overlap = set(part) & set(state)
+ if overlap:
+ raise ValueError(f"tensor(s) {sorted(overlap)[:3]} appear in two weight files")
+ state.update(part)
+ return state
+
+
+def _set_submodule(model, name, module):
+ parent_name, _, leaf = name.rpartition(".")
+ parent = model.get_submodule(parent_name) if parent_name else model
+ setattr(parent, leaf, module)
+
+
+def load_model_and_tokenizer(model_path, model_id, dtype, device, trust_remote_code):
+ from accelerate import init_empty_weights
+ from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, GenerationConfig
+
+ del model_id # the saved artifact is the only source of weights
+ # The artifact carries no modeling code (int8_export strips auto_map); the
+ # architecture is native to transformers, so remote code is never used.
+ del trust_remote_code
+ torch_dtype = _resolve_dtype(dtype)
+ qcfg = _read_quant_config(model_path)
+ state = _read_state(model_path)
+
+ config = AutoConfig.from_pretrained(
+ model_path, trust_remote_code=False, local_files_only=True
+ )
+ if getattr(config, "quantization_config", None):
+ raise ValueError("config.json carries a quantization_config; not a SmoothQuant int8 artifact")
+
+ # Parameters on meta (no fp16 copy of the linears is ever allocated); buffers
+ # such as the rotary inv_freq stay real.
+ with init_empty_weights(include_buffers=False):
+ model = AutoModelForCausalLM.from_config(
+ config,
+ torch_dtype=torch_dtype,
+ trust_remote_code=False,
+ attn_implementation="sdpa",
+ )
+
+ linear_cls = _w8a8_int8_linear_class()
+ modules = dict(model.named_modules())
+ quantized = []
+ for key in sorted(state):
+ if not key.endswith(SCALE_SUFFIX):
+ continue
+ name = key[: -len(SCALE_SUFFIX)]
+ weight = state.get(f"{name}.weight")
+ if weight is None or weight.dtype != torch.int8 or weight.dim() != 2:
+ raise ValueError(f"{key} has no int8 [out, in] weight next to it")
+ original = modules.get(name)
+ if not isinstance(original, nn.Linear):
+ raise ValueError(f"{name} is not an nn.Linear in {type(model).__name__}")
+ out_features, in_features = weight.shape
+ if (original.out_features, original.in_features) != (out_features, in_features):
+ raise ValueError(f"{name}: saved shape {tuple(weight.shape)} does not match the config")
+ if tuple(state[key].shape) not in ((out_features, 1), (out_features,)):
+ raise ValueError(f"{key}: expected [{out_features}, 1], got {tuple(state[key].shape)}")
+ state[key] = state[key].reshape(out_features, 1)
+ # The GEMM reads B as [N, K] with K contiguous; make sure that holds.
+ state[f"{name}.weight"] = weight.contiguous()
+ has_bias = f"{name}.bias" in state
+ if has_bias != (original.bias is not None):
+ raise ValueError(f"{name}: bias presence differs between artifact and config")
+ _set_submodule(model, name, linear_cls(in_features, out_features, has_bias, torch_dtype))
+ quantized.append(name)
+ if not quantized:
+ raise ValueError("no int8 weight + weight_scale pairs in the artifact")
+ expected = qcfg.get("num_quantized_linears")
+ if isinstance(expected, int) and expected != len(quantized):
+ raise ValueError(
+ f"{QUANT_CONFIG_NAME} lists {expected} quantized linears, artifact has {len(quantized)}"
+ )
+
+ for key, tensor in list(state.items()):
+ if tensor.is_floating_point() and tensor.dtype != torch_dtype:
+ state[key] = tensor.to(torch_dtype)
+
+ tied = bool(getattr(config, "tie_word_embeddings", False))
+ result = model.load_state_dict(state, strict=False, assign=True)
+ missing = set(result.missing_keys)
+ if tied:
+ missing.discard("lm_head.weight")
+ if missing or result.unexpected_keys:
+ raise ValueError(
+ f"artifact does not match the model: missing={sorted(missing)[:8]} "
+ f"unexpected={sorted(result.unexpected_keys)[:8]}"
+ )
+ if tied:
+ model.tie_weights()
+ leftover = [n for n, t in list(model.named_parameters()) + list(model.named_buffers()) if t.is_meta]
+ if leftover:
+ raise ValueError(f"tensors left uninitialised after load: {leftover[:8]}")
+
+ try:
+ model.generation_config = GenerationConfig.from_pretrained(
+ model_path, local_files_only=True
+ )
+ except OSError:
+ pass
+ model.to(device)
+ model.eval()
+ attn_impl = getattr(model.config, "_attn_implementation", None)
+ if attn_impl != "sdpa":
+ raise RuntimeError(f"expected SDPA attention for prefill, got {attn_impl!r}")
+
+ tokenizer = AutoTokenizer.from_pretrained(
+ model_path, trust_remote_code=False, local_files_only=True
+ )
+ return model, tokenizer
diff --git a/smoothquant/fake_quant.py b/smoothquant/fake_quant.py
index 24b62b5..bb7c175 100644
--- a/smoothquant/fake_quant.py
+++ b/smoothquant/fake_quant.py
@@ -13,6 +13,23 @@ def quantize_weight_per_channel_absmax(w, n_bits=8):
return w
+@torch.no_grad()
+def quantize_weight_per_channel_absmax_int8(w, n_bits=8):
+ """Same math as quantize_weight_per_channel_absmax, but also returns the integers.
+
+ Returns (q, scales, w_dq): q is int8 [out, in], scales is [out, 1] in w.dtype, and
+ w_dq = q * scales is bit-identical to what quantize_weight_per_channel_absmax
+ computes. w itself is not modified.
+ """
+ assert n_bits == 8, "int8 storage only"
+ scales = w.abs().max(dim=-1, keepdim=True)[0]
+ q_max = 2 ** (n_bits - 1) - 1
+ scales.clamp_(min=1e-5).div_(q_max)
+ q = w.div(scales).round_().clamp_(-q_max, q_max)
+ w_dq = q.mul(scales)
+ return q.to(torch.int8), scales, w_dq
+
+
@torch.no_grad()
def quantize_weight_per_tensor_absmax(w, n_bits=8):
# w: (out_features, in_features)
@@ -109,8 +126,15 @@ class W8A8Linear(nn.Module):
@staticmethod
def from_float(
- module, weight_quant="per_channel", act_quant="per_token", quantize_output=False
+ module,
+ weight_quant="per_channel",
+ act_quant="per_token",
+ quantize_output=False,
+ keep_int8=False,
):
+ """keep_int8=True also keeps the int8 weight and its per-channel scale as
+ buffers (weight_int8 [out, in], weight_scale [out, 1]) so the model can be
+ saved as real int8. The floating-point weight used in forward is unchanged."""
assert isinstance(module, torch.nn.Linear)
new_module = W8A8Linear(
module.in_features,
@@ -119,7 +143,16 @@ class W8A8Linear(nn.Module):
act_quant=act_quant,
quantize_output=quantize_output,
)
- if weight_quant == "per_channel":
+ if weight_quant == "per_channel" and keep_int8:
+ q, scales, w_dq = quantize_weight_per_channel_absmax_int8(
+ module.weight, n_bits=8
+ )
+ new_module.weight = w_dq
+ new_module.register_buffer("weight_int8", q)
+ new_module.register_buffer("weight_scale", scales)
+ elif keep_int8:
+ raise ValueError("keep_int8 is only supported with per_channel weights")
+ elif weight_quant == "per_channel":
new_module.weight = quantize_weight_per_channel_absmax(
module.weight, n_bits=8
) # use 8-bit integer for weight
@@ -310,16 +343,60 @@ def quantize_falcon(
return model
+def quantize_phi3(
+ model,
+ weight_quant="per_channel",
+ act_quant="per_token",
+ quantize_bmm_input=False,
+ keep_int8=False,
+):
+ """Phi-3: the Llama-like scheme on fused linears (qkv_proj, gate_up_proj)."""
+ # Phi3SdpaAttention / Phi3FlashAttention2 subclass Phi3Attention.
+ from transformers.models.phi3.modeling_phi3 import Phi3Attention, Phi3MLP
+
+ from_float = partial(
+ W8A8Linear.from_float,
+ weight_quant=weight_quant,
+ act_quant=act_quant,
+ keep_int8=keep_int8,
+ )
+ for name, m in model.model.named_modules():
+ if isinstance(m, Phi3MLP):
+ m.gate_up_proj = from_float(m.gate_up_proj)
+ m.down_proj = from_float(m.down_proj)
+ elif isinstance(m, Phi3Attention):
+ # q/k/v are one linear; quantize_output simulates int8 BMM inputs as the
+ # Llama path does for q_proj/k_proj/v_proj.
+ m.qkv_proj = from_float(m.qkv_proj, quantize_output=quantize_bmm_input)
+ m.o_proj = from_float(m.o_proj)
+ return model
+
+
def quantize_model(
- model, weight_quant="per_channel", act_quant="per_token", quantize_bmm_input=False
+ model,
+ weight_quant="per_channel",
+ act_quant="per_token",
+ quantize_bmm_input=False,
+ keep_int8=False,
):
from transformers.models.opt.modeling_opt import OPTPreTrainedModel
from transformers.models.llama.modeling_llama import LlamaPreTrainedModel
from transformers.models.mistral.modeling_mistral import MistralPreTrainedModel
from transformers.models.mixtral.modeling_mixtral import MixtralPreTrainedModel
from transformers.models.falcon.modeling_falcon import FalconPreTrainedModel
+ from transformers.models.phi3.modeling_phi3 import Phi3PreTrainedModel
- if isinstance(model, OPTPreTrainedModel):
+ if keep_int8 and not isinstance(model, Phi3PreTrainedModel):
+ raise ValueError("keep_int8 is only implemented for the Phi-3 path")
+ if isinstance(model, Phi3PreTrainedModel):
+ return quantize_phi3(
+ model,
+ weight_quant=weight_quant,
+ act_quant=act_quant,
+ quantize_bmm_input=quantize_bmm_input,
+ keep_int8=keep_int8,
+ )
+ elif isinstance(model, OPTPreTrainedModel):
return quantize_opt(
model,
weight_quant=weight_quant,
diff --git a/smoothquant/int8_export.py b/smoothquant/int8_export.py
new file mode 100644
index 0000000..7d1f5bf
--- /dev/null
+++ b/smoothquant/int8_export.py
@@ -0,0 +1,138 @@
+"""Save a smoothed, W8A8-quantized model as real int8 (the ``w8a8_int8`` format).
+
+quantize_model(..., keep_int8=True) leaves W8A8Linear modules that carry the
+int8 weight and its per-output-channel scale next to the floating-point weight
+they use in forward. This writes:
+
+- ``<linear>.weight`` int8 [out, in]
+- ``<linear>.weight_scale`` fp16 [out, 1] (w_fp16 = weight * weight_scale)
+- ``<linear>.bias`` fp16 [out] (unchanged by smoothing; when present)
+- every other tensor (smoothed norms, embeddings, final norm, an untied lm_head)
+ in fp16
+- no ``lm_head.weight`` when it is tied to the input embeddings
+
+plus config.json (no ``quantization_config`` and no ``auto_map``: the artifact
+carries no remote code and loads with the native transformers class),
+generation_config.json, the tokenizer, and ``quantization_config.json``
+describing the scheme. Activations are quantized per token, dynamically, at
+runtime; nothing about them is stored.
+"""
+import json
+import os
+
+import torch
+
+from smoothquant.fake_quant import W8A8Linear
+
+QUANT_CONFIG_NAME = "quantization_config.json"
+WEIGHTS_NAME = "model.safetensors"
+CONFIG_NAME = "config.json"
+
+
+@torch.no_grad()
+def w8a8_int8_state_dict(model, dtype=torch.float16):
+ """Return (state_dict, quantized module names) for the int8 artifact."""
+ int8_modules = {}
+ for name, module in model.named_modules():
+ if isinstance(module, W8A8Linear):
+ if getattr(module, "weight_int8", None) is None:
+ raise ValueError(
+ f"{name} was quantized without keep_int8=True; cannot save int8"
+ )
+ if getattr(module, "output_quant_name", "None") != "None":
+ raise ValueError(
+ f"{name} quantizes its output (quantize_bmm_input=True); "
+ "the w8a8_int8 format cannot represent that"
+ )
+ int8_modules[name] = module
+
+ state = {}
+ for key, tensor in model.state_dict().items():
+ owner = key.rpartition(".")[0]
+ if owner in int8_modules:
+ continue
+ tensor = tensor.detach()
+ if tensor.is_floating_point():
+ tensor = tensor.to(dtype)
+ state[key] = tensor.to("cpu", copy=True).contiguous()
+
+ for name, module in int8_modules.items():
+ weight = module.weight_int8.detach()
+ scale = module.weight_scale.detach()
+ out_features, in_features = module.out_features, module.in_features
+ if weight.dtype != torch.int8 or tuple(weight.shape) != (out_features, in_features):
+ raise ValueError(f"{name}: bad int8 weight {weight.dtype} {tuple(weight.shape)}")
+ if tuple(scale.shape) != (out_features, 1):
+ raise ValueError(f"{name}: bad weight_scale shape {tuple(scale.shape)}")
+ if not torch.isfinite(scale).all():
+ raise ValueError(f"{name}: non-finite weight_scale")
+ state[f"{name}.weight"] = weight.to("cpu", copy=True).contiguous()
+ state[f"{name}.weight_scale"] = scale.to("cpu", dtype=dtype, copy=True).contiguous()
+ if module.bias is not None:
+ bias = module.bias.detach().reshape(-1)
+ state[f"{name}.bias"] = bias.to("cpu", dtype=dtype, copy=True).contiguous()
+
+ for key, tensor in state.items():
+ if tensor.is_floating_point() and not torch.isfinite(tensor).all():
+ raise ValueError(f"{key}: non-finite values after casting to {dtype}")
+
+ if getattr(model.config, "tie_word_embeddings", False):
+ head = model.get_output_embeddings()
+ embed = model.get_input_embeddings()
+ if head is not None and head.weight is not embed.weight:
+ raise ValueError("config ties lm_head to embed_tokens but the weights differ")
+ for key in [k for k in state if k.endswith("lm_head.weight")]:
+ del state[key]
+ return state, sorted(int8_modules)
+
+
+def _strip_remote_code(output_dir):
+ """Drop auto_map: the artifact ships no modeling code and loads natively."""
+ path = os.path.join(output_dir, CONFIG_NAME)
+ with open(path) as handle:
+ config = json.load(handle)
+ if config.pop("auto_map", None) is None:
+ return False
+ with open(path, "w") as handle:
+ json.dump(config, handle, indent=2, sort_keys=True)
+ handle.write("\n")
+ return True
+
+
+@torch.no_grad()
+def save_w8a8_int8(model, tokenizer, output_dir, quant_config, dtype=torch.float16):
+ """Write the int8 artifact to output_dir. Returns a short summary dict."""
+ from safetensors.torch import save_file
+
+ os.makedirs(output_dir, exist_ok=True)
+ state, modules = w8a8_int8_state_dict(model, dtype=dtype)
+ if not modules:
+ raise ValueError("no W8A8Linear modules found; nothing was quantized")
+ save_file(state, os.path.join(output_dir, WEIGHTS_NAME), metadata={"format": "pt"})
+
+ config = model.config
+ if getattr(config, "quantization_config", None) is not None:
+ raise ValueError("model config already carries a quantization_config")
+ config.torch_dtype = dtype
+ config.save_pretrained(output_dir)
+ stripped = _strip_remote_code(output_dir)
+ generation_config = getattr(model, "generation_config", None)
+ if generation_config is not None:
+ generation_config.save_pretrained(output_dir)
+ if tokenizer is not None:
+ tokenizer.save_pretrained(output_dir)
+
+ record = dict(quant_config)
+ record["quantized_modules"] = modules
+ record["num_quantized_linears"] = len(modules)
+ with open(os.path.join(output_dir, QUANT_CONFIG_NAME), "w") as handle:
+ json.dump(record, handle, indent=2)
+ handle.write("\n")
+ int8_params = sum(t.numel() for t in state.values() if t.dtype == torch.int8)
+ return {
+ "output_dir": output_dir,
+ "num_quantized_linears": len(modules),
+ "int8_params": int8_params,
+ "tensors": len(state),
+ "auto_map_stripped": stripped,
+ }
diff --git a/smoothquant/smooth.py b/smoothquant/smooth.py
index c2a70e7..2ec010e 100644
--- a/smoothquant/smooth.py
+++ b/smoothquant/smooth.py
@@ -13,6 +13,7 @@ from transformers.models.mixtral.modeling_mixtral import (
MixtralRMSNorm,
)
from transformers.models.falcon.modeling_falcon import FalconDecoderLayer
+from transformers.models.phi3.modeling_phi3 import Phi3DecoderLayer, Phi3RMSNorm
@torch.no_grad()
@@ -49,7 +50,7 @@ def smooth_ln_fcs(ln, fcs, act_scales, alpha=0.5):
def smooth_ln_fcs_llama_like(ln, fcs, act_scales, alpha=0.5):
if not isinstance(fcs, list):
fcs = [fcs]
- assert isinstance(ln, (LlamaRMSNorm, MistralRMSNorm, MixtralRMSNorm))
+ assert isinstance(ln, (LlamaRMSNorm, MistralRMSNorm, MixtralRMSNorm, Phi3RMSNorm))
for fc in fcs:
assert isinstance(fc, nn.Linear)
assert ln.weight.numel() == fc.in_features == act_scales.numel()
@@ -158,3 +159,20 @@ def smooth_lm(model, scales, alpha=0.5):
fcs_input_scales = scales[name + ".block_sparse_moe.gate"]
smooth_ln_fcs_llama_like(ffn_ln, fcs, fcs_input_scales, alpha)
+ elif isinstance(module, Phi3DecoderLayer):
+ # Phi-3 fuses q/k/v into one qkv_proj and gate/up into one gate_up_proj.
+ # Each fused linear is the only reader of its norm, so it takes the whole
+ # smoothing scale; the column max over its rows equals the max over the
+ # separate q/k/v (gate/up) weights the Llama branch concatenates.
+ smooth_ln_fcs_llama_like(
+ module.input_layernorm,
+ [module.self_attn.qkv_proj],
+ scales[name + ".self_attn.qkv_proj"],
+ alpha,
+ )
+ smooth_ln_fcs_llama_like(
+ module.post_attention_layernorm,
+ [module.mlp.gate_up_proj],
+ scales[name + ".mlp.gate_up_proj"],
+ alpha,
+ )
|