QNN sharded LLM export fails: ResolveDebugHandle is not the last edge pass (dep_table entry overwritten by SplitGraph registration)

Ouverte
  • Quelque chose ne va pas ?
#23,580 0 commentaires 0 réactions 1 personne assignée Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

@psiddh y travaille déjà.

Depuis le 8/10/2026.

  • #23581 par @SabYic — ouverte

Évaluation

Cette issue n'a pas encore été évaluée.

Description

bug module: llm module: qnn triaged
🐛 Describe the bug

Exporting any sharded LLM through the QNN static-llama flow (num_sharding > 1, i.e. every model ≥ ~3B such as llama3_2-3b_instruct, or a Qwen3-4B registered locally) fails with:

AssertionError: Please ensure ResolveDebugHandle is the last executed edge pass.
  File "backends/qualcomm/_passes/qnn_pass_manager.py", line 432, in _validate_edge_passes

Non-sharded models (Qwen3 0.6B/1.7B, etc.) export fine, which is probably why CI has not caught this.

Root cause

When sharding is enabled, the dynamically created SplitGraph pass is registered like this (3 call sites):

dep_table[SplitGraph] = [FoldQDQ]
dep_table[TagQuantIO] = [SplitGraph]   # <-- overwrites TagQuantIO: [LayoutTransform]
  • examples/qualcomm/oss_scripts/llama/wrappers/llm_wrappers.py (two sites)
  • examples/models/llama/export_llama_lib.py (_export_llama, QNN branch)

The assignment replaces the pre-existing entry TagQuantIO: [LayoutTransform] from get_passes_dependency_for_capture_program(), silently dropping the LayoutTransform → TagQuantIO constraint. With that edge gone, the topological solver schedules TagQuantIO (and transitively ResolveDebugHandle, whose only constraint is "after TagQuantIO") too early, so ResolveDebugHandle is no longer last and the (recently added) _validate_edge_passes assertion fires.

Broken schedule tail: ... → ResolveDebugHandle → AnnotateGetAttr → InsertCastForFpActQuantizedWeight

Minimal repro (no device, no weights, runs in seconds):

from torch.fx.passes.infra.pass_manager import this_before_that_pass_constraint
from executorch.backends.qualcomm._passes import FoldQDQ, TagQuantIO
from executorch.backends.qualcomm._passes.qnn_pass_manager import get_qnn_pass_manager_cls
from executorch.extension.llm.custom_ops import model_sharding

cls = get_qnn_pass_manager_cls()
pm = cls()
passes_job = cls.get_capture_program_passes()
dep_table = cls.get_passes_dependency_for_capture_program()
SplitGraph, setting = model_sharding.get_split_graph_pass(36, shares=4, pattern=r"^layers\.(\d+)")
passes_job[SplitGraph] = setting
dep_table[SplitGraph] = [FoldQDQ]
dep_table[TagQuantIO] = [SplitGraph]          # verbatim from the call sites above

for that, these in dep_table.items():
    for this in these:
        pm.add_constraint(this_before_that_pass_constraint(this, that))
for p in passes_job:
    pm.add_pass(p)
pm.solve_constraints()
order = [getattr(p, "__name__", str(p)) for p in pm.passes]
assert order[-1] == "ResolveDebugHandle", f"tail: {order[-3:]}"   # fails on main

End-to-end repro: python examples/qualcomm/oss_scripts/llama/llama.py --decoder_model llama3_2-3b_instruct ... (any sharded model).

Proposed fix (PR to follow):

  1. Call sites: append instead of overwrite — dep_table.setdefault(TagQuantIO, []).append(SplitGraph).
  2. Structural guarantee in QnnPassManager.get_to_edge_transform_passes: constrain ResolveDebugHandle after every other pass in passes_job, so dynamically registered passes can never be scheduled after it. This keeps the invariant that _validate_edge_passes asserts, instead of relying on call sites to remember it.

Verified: with the fix, a sharded Qwen3-4B (16a4w, 4 shards) exports and runs on SM8650 (113.8 tok/s prefill, 7.24 tok/s decode), and TestQNNQuantizedOperator.test_qnn_backend_abs --enable_x86_64 passes.

Found and analyzed with AI assistance (Claude Code).

Versions
  • executorch: main @ 5e21c13cc9e34fa2922d5708ba5962e3e365afdb (2026-10-05)
  • QNN SDK: 2.37.0.250724, SoC: SM8650 (HTP v75)
  • torch 2.14.0+cpu, Python 3.11, Ubuntu (Linux 5.15)

cc @cbilgin @psiddh @larryliu0820 @mergennachin @cccclai @helunwencser @jackzhxng @digantdesai

Langage dominant
Python
Étoiles
5.1k
Forks
1.2k
Merge moyen
2 j 6 h
PR mergées (30 j)
692

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de pytorch/executorch

Toutes les issues de pytorch/executorch

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.