LFM2/LFM2.5: short-conv state is not cleared when a new sequence starts, so output after reset() depends on the previous prompt

未关闭

#23,262 创建于 2026年9月29日 0 条评论 0 个 reaction 已指派 0 人在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度: 4/5
预计耗时: 3-5 天
新手友好度: 68/100

Issue 类型: 缺陷
描述清晰度: 描述清楚
活跃度: 活跃
技术栈: python, pytorch
领域: machine-learning, testing-qa

调研方向

Start with examples/models/lfm2/short_conv.py, especially ShortConv.forward, ShortConvBlock.forward, and reset_cache(), then compare the reset behavior in examples/models/llama/attention.py around _maybe_reset_state. Run the supplied module and TextLLMRunner reproductions, and add a regression test like test_gated_deltanet_resets_state_on_new_sequence. Done means a new sequence after reset matches a freshly loaded model without changing continuation behavior.

由索引模型根据 Issue 内容生成。

描述

🐛 Describe the bug

Summary

In exported LFM2 and LFM2.5 models, ShortConv.conv_state (a mutable buffer holding the last two conv inputs) is never cleared when a new sequence starts at input_pos = 0. After TextLLMRunner.reset() (Python), LlmModule.resetContext() (Android) or TextLLMRunner::warmup(), the next conversation starts from the previous conversation's convolution state. Continuing a conversation at the current position is not affected.

Expected: after reset(), a prompt gives the same output as on a freshly loaded model.
Actual: the output depends on what ran before the reset.

To reproduce

Environment: executorch==1.5.1 and torch==2.14.0 from pip.
Model: a public export, lfm_2_5_350m_xnnpack_8da4w.pte (sha256 e8c38e0d933932b02dddbcf92df813045899d0f71bf5346dbceecf5c4078f05d), and its tokenizer.json (sha256 2221c71b5dce048a8abae62843b92bd7deec13cc153f95fa6e3327a47b79a7da).

1. Module level (python module_repro.py lfm_2_5_350m_xnnpack_8da4w.pte)

import sys
import torch
from executorch.extension.llm.custom_ops import custom_ops  # noqa: F401
from executorch.kernels import quantized  # noqa: F401
from executorch.extension.pybindings.portable_lib import _load_for_executorch

PTE = sys.argv[1]  # e.g. lfm_2_5_350m_xnnpack_8da4w.pte

def run(module, tokens):
    for pos, tok in enumerate(tokens):
        logits = module.forward((torch.tensor([[tok]]), torch.tensor([pos])))[0]
    return logits.reshape(-1).float()

A = [1, 6423, 1098, 4123, 872, 3301, 911]  # an earlier, unrelated sequence
B = [1, 1715, 5902, 3314, 7310]            # a new sequence, positions restart at 0

fresh = run(_load_for_executorch(PTE), B)
reused = _load_for_executorch(PTE)
run(reused, A)
after_a = run(reused, B)

print("max |fresh - after_a| =", (fresh - after_a).abs().max().item())   # expected 0.0
print("argmax fresh / after_a:", fresh.argmax().item(), after_a.argmax().item())

Output:

max |fresh - after_a| = 16.080059051513672
argmax fresh / after_a: 7 3466

Controls: two fresh loads give 0.0, and the same script on an attention-only model (a Qwen3-1.7B XNNPACK export) gives 0.0.

2. Through TextLLMRunner, greedy, with reset() (python runner_repro.py lfm_2_5_350m_xnnpack_8da4w.pte tokenizer.json)

import sys
from executorch.extension.llm.custom_ops import custom_ops  # noqa: F401
from executorch.kernels import quantized  # noqa: F401
from executorch.extension.llm.runner import GenerationConfig, TextLLMRunner

PTE, TOKENIZER = sys.argv[1], sys.argv[2]

def chat(q):
    return f"<|startoftext|><|im_start|>user\n{q}<|im_end|>\n<|im_start|>assistant\n"

def generate(runner, prompt):
    out = []
    runner.generate(prompt, GenerationConfig(echo=False, max_new_tokens=64, temperature=0.0), out.append)
    return "".join(out)

fresh = generate(TextLLMRunner(PTE, TOKENIZER), chat("List five uses for a paperclip."))

runner = TextLLMRunner(PTE, TOKENIZER)
generate(runner, chat("Write a short poem about the ocean at night."))
runner.reset()  # documented as "Reset the runner state and KV cache"
after_reset = generate(runner, chat("List five uses for a paperclip."))

print("identical:", fresh == after_reset)          # expected True
print("fresh      :", repr(fresh[:60]))
print("after reset:", repr(after_reset[:60]))

Last three lines of output (the runner also streams each generation):

identical: False
fresh      : 'Sure! Here are five practical uses for a paperclip:\n\n1. **St'
after reset: 'Here are five uses for a paperclip:\n\n1. **Home Decor**: Pape'

Scope

Each export was checked on 4 random sequence pairs (harness and raw output below). "Leaks" means every pair differed while fresh vs fresh stayed at 0.0.

Export	Backend, dtype	Result
LFM2-350M, exported with the README's first example	XNNPACK 8da4w	leaks (4/4)
LFM2.5-350M, same recipe with `lfm2_5_350m` config (and `lfm2_xnnpack_fp32.yaml` for fp32)	XNNPACK 8da4w and fp32	leaks (4/4 each)
Software Mansion LFM2.5-350M and 1.2B	XNNPACK 8da4w	leaks (4/4 each)
Software Mansion LFM2.5-1.2B	XNNPACK fp16	leaks (4/4)
`younghan-meta/LFM2.5-ExecuTorch-MLX` 350M	MLX 4-bit	leaks (4/4)
Qwen3-1.7B (attention only, control)	XNNPACK	no leak (0/4)

Reproduced on macOS arm64; the Software Mansion 350M file also leaks (4/4) on Linux aarch64 (python:3.12-slim, pip install executorch==1.5.1 torch==2.14.*). A short, confident answer can come out unchanged because the logits shift without always changing the argmax; open-ended prompts diverge.

Root cause

Permalinks at 2f78245. short_conv.py is identical in v1.4.0 and v1.5.1, and so are the TextLLMRunner::reset() and warmup() bodies.

conv_state is registered at short_conv.py#L42, then prepended and overwritten on every call at #L66-L75. The comment at #L63 assumes prefill starts on an empty cache.
ShortConvBlock.forward receives the attention options that carry input_pos but ignores them (#L104). reset_cache() (#L85) is never called outside eager code.
TextLLMRunner::reset() (text_llm_runner.cpp#L354) resets stats, pos_ and the pending prefill token only, and warmup() (#L331) calls it after generating. The Python binding documents reset() as "Reset the runner state and KV cache" (pybindings.cpp#L735).

Verified fix

Qwen3.5's GatedDeltaNet already clears its state when input_pos[0] == 0 (_maybe_reset_state, attention.py#L801-L811). Applying the same pattern to ShortConv and re-exporting with the same recipe removes the leak:

def conv_forward(self, x, input_pos=None):              # ShortConv
    if input_pos is not None:
        self.conv_state.mul_(1.0 - (input_pos[0] == 0).to(self.conv_state.dtype))
    return original_conv_forward(self, x)

def block_forward(self, x, freqs_cos=None, freqs_sin=None, _unused_attn_options=None):   # ShortConvBlock
    input_pos = (_unused_attn_options or {}).get("input_pos")
    h = x + self.conv.forward(self.attention_norm(x), input_pos)
    return h + self.feed_forward(self.ffn_norm(h)), None

Check	Unpatched	Patched
Repro 1, max logit difference	15.55	0.0
Repro 2	`identical: False`	`identical: True`
Random pairs leaking	4/4	0/4
Fresh sequence, patched vs unpatched logits		identical (0.0)

(LFM2.5-350M fp32 export, same recipe. The patch is a no-op while input_pos[0] != 0 and clears conv_state whenever a sequence starts at position 0, with or without reset().)

Suggested next steps

Apply the patch above and add a regression test like Qwen3.5's test_gated_deltanet_resets_state_on_new_sequence. If input_pos=None is supported, reset state on that path too, as _maybe_reset_state does.
Files already exported keep the leak until re-exported, so clearing mutable buffers in reset() may also be worth considering.

I can open a PR with the model-side change and a test.

#19562 (open) adds a long-lived LFM2.5-350M MLX formatter whose helper calls runner.reset() so that "each request is independent"; with this bug the requests are not.
#21858 (LFM2.5 degenerate through TextLLMRunner on the first call) looks like a different symptom; I have not established whether the two are connected.

Harness and raw output (all exports)

"""Quadruple check of the LFM2 state leak. Module-level: does a new sequence at input_pos 0 on a reused module give
the same logits as the same sequence on a freshly loaded module? Several random sequence pairs, several lengths.

  python harness.py <file.pte> [vocab_cap]
"""
import random, sys, hashlib
import torch
from executorch.extension.llm.custom_ops import custom_ops  # noqa: F401
from executorch.kernels import quantized  # noqa: F401
from executorch.extension.pybindings.portable_lib import _load_for_executorch

pte = sys.argv[1]
cap = int(sys.argv[2]) if len(sys.argv) > 2 else 8000
sha = hashlib.sha256(open(pte, "rb").read()).hexdigest()[:16]

def run(m, toks):
    out = None
    for i, t in enumerate(toks):
        out = m.forward((torch.tensor([[t]], dtype=torch.long), torch.tensor([i], dtype=torch.long)))[0]
    return out.reshape(-1).float()

def kl(a, b):
    p, q = torch.log_softmax(a, -1), torch.log_softmax(b, -1)
    return float((p.exp() * (p - q)).sum())

m = _load_for_executorch(pte)
print(f"{pte.split('/')[-1]} sha256:{sha} methods={[x for x in m.method_names() if x == 'forward' or 'reset' in x]}")
rng = random.Random(0)
rows = []
for trial, (la, lb) in enumerate([(1, 4), (6, 5), (12, 8), (3, 3)]):
    A = [1] + [rng.randrange(100, cap) for _ in range(la)]
    B = [1] + [rng.randrange(100, cap) for _ in range(lb)]
    fresh = run(_load_for_executorch(pte), B)
    fresh2 = run(_load_for_executorch(pte), B)
    m2 = _load_for_executorch(pte); run(m2, A); after = run(m2, B)
    m3 = _load_for_executorch(pte); run(m3, B); again = run(m3, B)     # same sequence twice on one module
    rows.append((trial, la, lb, float((fresh - fresh2).abs().max()), float((fresh - after).abs().max()),
                 int(fresh.argmax()) != int(after.argmax()), kl(fresh, after), float((fresh - again).abs().max())))
for r in rows:
    print("  trial %d |A|=%-2d |B|=%-2d  control %.6f | A-then-B max %.4f argmax-changed %-5s KL %.4f | B-then-B max %.4f" % r)
leak = sum(r[4] > 1e-3 for r in rows)
print(f"  => {leak}/{len(rows)} trials leak; control max {max(r[3] for r in rows):.6f}")

# harness.py: 4 random sequence pairs per file; executorch 1.5.1 (pip), torch 2.14.0, macOS 26.5.2 arm64
swm_350m_8da4w.pte sha256:e8c38e0d933932b0 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 16.8667 argmax-changed True  KL 7.7509 | B-then-B max 19.7846
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 18.1271 argmax-changed True  KL 5.0337 | B-then-B max 14.6948
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 15.9862 argmax-changed True  KL 5.9448 | B-then-B max 17.6679
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 14.2549 argmax-changed True  KL 2.6568 | B-then-B max 15.3156
  => 4/4 trials leak; control max 0.000000
swm_1_2b_8da4w.pte sha256:69b9d1d7f7d576e1 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 6.5088 argmax-changed True  KL 1.1331 | B-then-B max 7.9537
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 10.9939 argmax-changed True  KL 2.5087 | B-then-B max 9.8512
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 4.0695 argmax-changed False KL 0.2427 | B-then-B max 8.2273
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 6.7117 argmax-changed True  KL 1.8102 | B-then-B max 9.4184
  => 4/4 trials leak; control max 0.000000
swm_1_2b_fp16.pte sha256:cd013f51d5f7f3a7 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 8.8141 argmax-changed True  KL 1.5633 | B-then-B max 9.9648
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 9.6172 argmax-changed True  KL 1.7002 | B-then-B max 7.9590
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 3.7124 argmax-changed False KL 0.2932 | B-then-B max 10.9331
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 5.7681 argmax-changed False KL 0.5080 | B-then-B max 6.1289
  => 4/4 trials leak; control max 0.000000
yh_350m_mlx_4w.pte sha256:a22641ff8364b813 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 19.9961 argmax-changed True  KL 5.9034 | B-then-B max 18.8750
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 18.4062 argmax-changed True  KL 5.4028 | B-then-B max 15.4062
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 16.9062 argmax-changed True  KL 8.4498 | B-then-B max 15.3125
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 16.5312 argmax-changed True  KL 3.9572 | B-then-B max 15.4570
  => 4/4 trials leak; control max 0.000000
fresh_lfm2_5_350m_8da4w.pte sha256:e7e1fa1c7f357243 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 17.0455 argmax-changed True  KL 7.5597 | B-then-B max 18.3524
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 14.5359 argmax-changed True  KL 3.8377 | B-then-B max 13.2261
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 16.6925 argmax-changed True  KL 6.4640 | B-then-B max 13.8776
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 15.6679 argmax-changed True  KL 3.1766 | B-then-B max 19.1684
  => 4/4 trials leak; control max 0.000000
fresh_lfm2_5_350m_fp32.pte sha256:ed48c2bacfc9c040 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 17.0677 argmax-changed True  KL 5.6892 | B-then-B max 21.2208
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 19.8826 argmax-changed False KL 4.5853 | B-then-B max 16.9365
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 19.6481 argmax-changed True  KL 8.9908 | B-then-B max 14.9544
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 11.9439 argmax-changed True  KL 2.5607 | B-then-B max 17.1218
  => 4/4 trials leak; control max 0.000000
patched_lfm2_5_350m_fp32.pte sha256:d46fe088f609b2ba methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  => 0/4 trials leak; control max 0.000000
model.pte sha256:074d87c3cba37ed8 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 0.0000 argmax-changed False KL 0.0000 | B-then-B max 0.0000
  => 0/4 trials leak; control max 0.000000

# Linux aarch64 (python:3.12-slim container, pip install executorch==1.5.1 torch==2.14.*)
Linux-6.8.0-117-generic-aarch64-with-glibc2.41 torch 2.14.0+cu130
swm_350m_8da4w.pte sha256:e8c38e0d933932b0 methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 16.8667 argmax-changed True  KL 7.7509 | B-then-B max 19.7846
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 18.1271 argmax-changed True  KL 5.0337 | B-then-B max 14.6948
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 16.0775 argmax-changed True  KL 5.9302 | B-then-B max 17.6679
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 14.2549 argmax-changed True  KL 2.6568 | B-then-B max 15.3156
  => 4/4 trials leak; control max 0.000000

# LFM2-350M exported with the README's first example command verbatim (lfm2_xnnpack_q8da4w.yaml, lfm2_350m)
lfm2_350m_8da4w.pte sha256:adf2915039c60deb methods=['forward']
  trial 0 |A|=1  |B|=4   control 0.000000 | A-then-B max 9.1079 argmax-changed True  KL 2.4930 | B-then-B max 7.7957
  trial 1 |A|=6  |B|=5   control 0.000000 | A-then-B max 11.9951 argmax-changed True  KL 3.9635 | B-then-B max 12.0744
  trial 2 |A|=12 |B|=8   control 0.000000 | A-then-B max 12.3841 argmax-changed True  KL 3.1355 | B-then-B max 12.4198
  trial 3 |A|=3  |B|=3   control 0.000000 | A-then-B max 11.2085 argmax-changed True  KL 3.0994 | B-then-B max 10.8808
  => 4/4 trials leak; control max 0.000000

Versions

executorch 1.5.1 (pip), torch 2.14.0, torchao 0.18.0, Python 3.12.13, macOS 26.5.2 arm64. The Software Mansion 350M result also reproduces on Linux aarch64 (python:3.12-slim container, pip install executorch==1.5.1 torch==2.14.*). Source checked at main 2f7824593d0d14f9d5d73faa540b681b6b2b507c, v1.5.1 and v1.4.0.

collect_env

Collecting environment information...
PyTorch version: 2.14.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 26.5.2 (arm64)
GCC version: Could not collect
Clang version: 21.0.0 (clang-2100.1.1.101)
CMake version: version 4.4.0
Libc version: N/A

Python version: 3.12.13 (main, Jun 23 2026, 15:44:24) [Clang 22.1.3 ] (64-bit runtime)
Python platform: macOS-26.5.2-arm64-arm-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: False
Caching allocator config: N/A

CPU:
Apple M5

Versions of relevant libraries:
[pip3] Could not collect
[conda] Could not collect
executorch             1.5.1
numpy                  2.5.3
pytorch-tokenizers     1.5.0
torch                  2.14.0
torchao                0.18.0

主要语言: Python
星标: 5k
派生: 1.2k
平均合并: 2 天 7 小时
30 天内合并 PR: 521

环境准备

阅读贡献指南

从这里开始

先读完整个 Issue，再读项目的贡献指南。
在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
Fork 仓库，在一个分支上完成修改。
提交 Pull Request，并在描述里引用这个 Issue 编号。

pytorch/executorch 的其他 Issue

Automate dependency updates with a scheduled Action
未关闭
- 哪里不对？
enhancement triaged

难度 2/5 半天新手友好度 68/100

pytorch/executorch#21640 · 2026年8月7日创建

维护者通常 1 天内回复
[Android] LlmGenerationConfig default temperature (0.8) overrides LlmModule temperature in the config generate overload
未关闭
- 哪里不对？
难度 4/5 3-5 天新手友好度 55/100

pytorch/executorch#23263 · 2026年9月29日创建

维护者通常 1 天内回复
XNNPACK never delegates slice_copy with stride != 1, but op-support.csv only excludes zero-dim/dynamic shapes
可能已有人在做 @JakeStevens 今天认领。未关闭
- 哪里不对？
module: xnnpack

pytorch/executorch#23258 · 已指派 1 人 · 2026年9月29日创建

维护者通常 1 天内回复
[RFC] ExecuTorch Persisting Device Specialized Delegate Artifacts
可能已有人在做 @JacobSzwejbka 于 1 天前认领。未关闭
- 哪里不对？
pytorch/executorch#23192 · 1 条评论 · 1 个 reaction · 已指派 1 人 · 2026年9月28日创建

维护者通常 1 天内回复
[QNN] Enable ConvTranspose + BatchNorm fusion after #23170
可能已有人在做 @psiddh 于 1 天前认领。未关闭
- 哪里不对？
module: qnn partner: qualcomm

pytorch/executorch#23185 · 1 条评论 · 1 个 reaction · 已指派 1 人 · 2026年9月28日创建

维护者通常 1 天内回复

查看 pytorch/executorch 的全部 Issue

相似的 Issue

Add support for cli completion in git bash for windows
未关闭
- 哪里不对？
customer-reported

难度 2/5 1-3 小时新手友好度 68/100

Azure/azure-cli#34150 · 1 条评论 · 2026年9月30日创建

维护者通常 1 天内回复
docs: fix typo in version selector CSS comment
未关闭
- 哪里不对？
community-request

难度 1/5 1 小时以内新手友好度 95/100

NVIDIA-NeMo/Curator#2464 · 1 条评论 · 2026年9月30日创建

维护者通常 1 天内回复
weblate-discover crashes with an unhandled FileNotFoundError when the directory does not exist
未关闭
- 哪里不对？
难度 2/5 1-3 小时新手友好度 88/100

WeblateOrg/translation-finder#1099 · 2026年9月30日创建

维护者通常 1 天内回复
xtask memory usage leaves out sections with names of 15 or more characters
未关闭
- 哪里不对？
难度 2/5 1-3 小时新手友好度 68/100

trezor/trezor-firmware#7997 · 2026年9月30日创建

维护者通常 2 天内回复
TypedTableBlock Duplicate changes the caption to literal undefined
未关闭
- 哪里不对？
难度 2/5 1-3 小时新手友好度 88/100

wagtail/wagtail#14662 · 2026年9月30日创建

维护者通常 1 天内回复

更多 Python Issue

LFM2/LFM2.5: short-conv state is not cleared when a new sequence starts, so output after reset() depends on the previous prompt

哪里不对？

评估

调研方向

描述

🐛 Describe the bug

Summary

To reproduce

Scope

Root cause

Verified fix

Suggested next steps

Related

Versions

环境准备

从这里开始

pytorch/executorch 的其他 Issue

哪里不对？

哪里不对？

哪里不对？

哪里不对？

哪里不对？

相似的 Issue

哪里不对？

哪里不对？

哪里不对？

哪里不对？

哪里不对？

把新 issue 发到你的邮箱