bump glm patch

.pre-commit-config.yaml

@@ -15,10 +15,15 @@ repos:
     rev: v6.0.0
     hooks:
       - id: trailing-whitespace
+        exclude: '\.patch$'
       - id: check-yaml
       - id: end-of-file-fixer
+        exclude: '\.patch$'
       - id: requirements-txt-fixer
       - id: double-quote-string-fixer
+        exclude: '\.patch$'
       - id: check-merge-conflict
+        exclude: '\.patch$'
       - id: mixed-line-ending
         args: ["--fix=lf"]
+        exclude: '\.patch$'

README.md

@@ -110,7 +110,7 @@ Recommended Runtime Environment:
 | torch        | >=2.0        | 2.8.0/2.11.0       |                    |
 | transformer-engine    | >=2.3       |  2.14.1    |                  |
 | apex |         | 0.1 | Optional |
-| megatron-core    |   >=0.16,<0.20    | 0.17.1      |                  |
+| megatron-core    |   >=0.16,<0.21    | 0.17.1      |                  |
 | flash-attn    |        | 2.8.3/3.0.0b1   |     Optional      |
 | transformers | >=4.33       | 4.57.6/5.8.1   |                    |
 | modelscope   | >=1.23       |             |                    |

README_zh.md

@@ -108,7 +108,7 @@ uv pip install -e . --torch-backend=auto
 | torch        | >=2.0        | 2.8.0/2.11.0       |                    |
 | transformer-engine    | >=2.3       |  2.14.1    |                  |
 | apex |       | 0.1 | 可选 |
-| megatron-core    |   >=0.16,<0.20    | 0.17.1      |                  |
+| megatron-core    |   >=0.16,<0.21    | 0.17.1      |                  |
 | flash-attn    |        | 2.8.3/3.0.0b1   |       可选           |
 | transformers | >=4.33       | 4.57.6/5.8.1   |                    |
 | modelscope   | >=1.23       |             |                    |

requirements.txt

@@ -1,4 +1,4 @@
-megatron-core>=0.16,<0.20
+megatron-core>=0.16,<0.21
 modelscope
 peft>=0.11,<0.21
 safetensors

src/mcore_bridge/config/model_config.py

@@ -250,6 +250,10 @@ class ModelConfig(TransformerConfig):
     csa_compress_rotary_base: float = 40000.0
     o_groups: int = 8
     o_lora_rank: int = 1024
+    # Canonical names consumed by TransformerConfig.__post_init__ since the main->dev sync (0.20+);
+    # `o_groups` / `o_lora_rank` remain the dev-era aliases and are reconciled against these.
+    output_projection_groups: int = 8
+    output_projection_lora_rank: int = 1024
     enable_hyper_connections: bool = False
     num_residual_streams: int = 4
     mhc_sinkhorn_iterations: int = 20

src/mcore_bridge/model/mm_gpts/glm5_next.py

@@ -1,5 +1,6 @@
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import copy
+import inspect
 import torch
 import torch.nn as nn
 from copy import deepcopy
@@ -34,6 +35,15 @@
 
 logger = get_logger()
 
+# Megatron 0.20 added a required `config` argument to select_pipeline_segment (the hybrid stack now
+# returns per-layer TransformerConfig objects instead of layer symbols); 0.18/0.19 take
+# `(main_pattern, pp_group, vp_stage, ...)`. GLM-5.3 itself is a dev-only feature -- a release without
+# the packaged patch is rejected earlier by require_glm5_hybrid() -- so only the 0.20 signature runs
+# today. Probe it anyway so a backport (or a release that grows the hybrid stack) cannot silently
+# turn this call site into a TypeError.
+_SELECT_PIPELINE_SEGMENT_TAKES_CONFIG = (
+    select_pipeline_segment is not None and 'config' in inspect.signature(select_pipeline_segment).parameters)
+
 # Probes for require_glm5_hybrid(): the hybrid-stack APIs this model is built on, and the
 # TransformerConfig fields that only the packaged patch provides.
 _HYBRID_API_PROBES = (('megatron.core.models.hybrid.hybrid_model', 'HybridModel'),
@@ -227,15 +237,17 @@ def __init__(self, config, transformer_layer_spec, pre_process=True, post_proces
         from megatron.core.process_groups_config import ProcessGroupCollection
         pg = ProcessGroupCollection.use_mpu_process_groups()
         pattern = self._resolve_hybrid_layer_pattern(config)
-        segment, offset = select_pipeline_segment(
-            parse_hybrid_pattern(pattern).main_pattern,
-            pg.pp,
-            vp_stage,
-            first_stage_layers=config.num_layers_in_first_pipeline_stage,
-            last_stage_layers=config.num_layers_in_last_pipeline_stage,
-            tp_group=pg.tp,
-            dp_cp_group=pg.dp_cp,
-        )
+        segment_kwargs = {
+            'first_stage_layers': config.num_layers_in_first_pipeline_stage,
+            'last_stage_layers': config.num_layers_in_last_pipeline_stage,
+            'tp_group': pg.tp,
+            'dp_cp_group': pg.dp_cp,
+        }
+        main_pattern = parse_hybrid_pattern(pattern).main_pattern
+        if _SELECT_PIPELINE_SEGMENT_TAKES_CONFIG:
+            segment, offset = select_pipeline_segment(main_pattern, config, pg.pp, vp_stage, **segment_kwargs)
+        else:
+            segment, offset = select_pipeline_segment(main_pattern, pg.pp, vp_stage, **segment_kwargs)
         if offset % 2 or len(segment) % 2:
             raise ValueError(
                 f'Pipeline stage boundaries of the current model must fall on complete attention+FFN blocks, '