Skip to content

Commit 2116952

Browse files
authored
Merge branch 'main' into fix/multimodal-add-document-default-modes
2 parents b7a3fea + 9f3fe8c commit 2116952

66 files changed

Lines changed: 3818 additions & 1150 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

‎.github/oncall_schedule.json‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,4 @@
11
[
2-
{
3-
"user": "ilml",
4-
"date": "2026-07-29"
5-
},
62
{
73
"user": "janEbert",
84
"date": "2026-08-05"
@@ -46,5 +42,9 @@
4642
{
4743
"user": "ilml",
4844
"date": "2026-10-14"
45+
},
46+
{
47+
"user": "janEbert",
48+
"date": "2026-10-21"
4949
}
5050
]

‎.gitlab-ci.yml‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -283,7 +283,7 @@ variables:
283283
description: Apply proposed Linear issue opens, updates, and closes
284284

285285
# CI wide variables
286-
NEMO_CI_TRIAGE_CONFIG: .gitlab/nemo-ci-triage.yml
286+
CERNO_CONFIG: .gitlab/cerno.yml
287287
CI_MCORE_LTS_IMAGE: ${GITLAB_ENDPOINT}:5005/adlr/megatron-lm/mcore_ci_lts
288288
CI_MCORE_DEV_IMAGE: ${GITLAB_ENDPOINT}:5005/adlr/megatron-lm/mcore_ci_dev
289289
CI_NEMO_IMAGE: ${GITLAB_ENDPOINT}:5005/adlr/megatron-lm/nemo_ci
Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,4 @@
1-
# Megatron-LM configuration for nemo-ci-triage.
1+
# Megatron-LM configuration for Cerno.
22

33
gitlab:
44
project_id: 19378

‎.gitlab/scripts/build.sh‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -50,7 +50,7 @@ fi
5050

5151
if [[ "$FILE" == "Dockerfile.linting" ]]; then
5252
ADDITIONAL_PARAMS+=("--build-arg CI_SERVER_URL=${CI_SERVER_URL}")
53-
ADDITIONAL_PARAMS+=("--secret id=NEMO_CI_TRIAGE_TOKEN,env=PAT")
53+
ADDITIONAL_PARAMS+=("--secret id=CERNO_TOKEN,env=PAT")
5454
fi
5555

5656
echo $(git rev-parse HEAD)

‎.gitlab/stages/06.triage.yml‎

Lines changed: 8 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -25,12 +25,12 @@ triage:linear_reconcile:
2525
artifacts: true
2626
script:
2727
- >-
28-
nemo-ci-linear status
29-
--config "${NEMO_CI_TRIAGE_CONFIG}"
28+
cerno-linear status
29+
--config "${CERNO_CONFIG}"
3030
--build-module-regex '^megatron-lm$'
3131
--output linear_status_report.json
3232
- >-
33-
nemo-ci-linear reconcile
33+
cerno-linear reconcile
3434
--failure-buckets failure_buckets.json
3535
--linear-report linear_status_report.json
3636
--pipeline-summaries pipeline_summaries.json
@@ -50,8 +50,8 @@ triage:linear_write:
5050
allow_failure: true
5151
script:
5252
- >-
53-
nemo-ci-linear write
54-
--config "${NEMO_CI_TRIAGE_CONFIG}"
53+
cerno-linear write
54+
--config "${CERNO_CONFIG}"
5555
--plan linear_action_plan.json
5656
--output linear_action_plan_post.json
5757
artifacts:
@@ -79,7 +79,7 @@ triage:slack_linear_followup:
7979
allow_failure: true
8080
script:
8181
- >-
82-
nemo-ci-notify
82+
cerno-notify
8383
--pipeline-summary slack_notification.json
8484
--linear-plan linear_action_plan_post.json
8585
--slack-bot-token "${MCORE_SLACK_BOT_TOKEN:-${ALERTMANAGER_TOKEN}}"
@@ -89,8 +89,8 @@ triage:slack_linear_followup:
8989
if [[ -z "${THREAD_TIMESTAMP}" ]]; then
9090
echo "No Slack thread timestamp; skipping detailed triage follow-ups."
9191
else
92-
nemo-ci-notify \
93-
--config "${NEMO_CI_TRIAGE_CONFIG}" \
92+
cerno-notify \
93+
--config "${CERNO_CONFIG}" \
9494
--module megatron_lm \
9595
--only-followup \
9696
--thread-ts "${THREAD_TIMESTAMP}" \

‎docker/Dockerfile.ci.dev‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -82,7 +82,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
8282
EOF
8383

8484
# Install DeepEP
85-
ARG DEEPEP_COMMIT=17cfb817bccec3a9c247013360cc550c2bac441e
85+
ARG DEEPEP_COMMIT=de0dd1185142c727b9c118aeb4cc4702e34deeba
8686
ENV DEEPEP_COMMIT=$DEEPEP_COMMIT
8787
ENV HYBRID_EP_MULTINODE=1
8888
ENV RDMA_CORE_HOME=/opt/rdma-core/build

‎docker/Dockerfile.linting‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -24,10 +24,10 @@ RUN --mount=type=secret,id=JET_INDEX_URLS \
2424

2525
# Keep this in the internal-only stage so public CI has no internal service dependency.
2626
ARG CI_SERVER_URL
27-
ARG NEMO_CI_TRIAGE_COMMIT=6e24e567ae2855e8acad5e2f780c7c26c195ab46
28-
RUN --mount=type=secret,id=NEMO_CI_TRIAGE_TOKEN \
27+
ARG CERNO_COMMIT=5a5fb5360e67f8f09d189871bbc0d768c09c43fa
28+
RUN --mount=type=secret,id=CERNO_TOKEN \
2929
GIT_CONFIG_COUNT=1 \
3030
GIT_CONFIG_KEY_0=http.extraHeader \
31-
GIT_CONFIG_VALUE_0="Authorization: Basic $(printf 'oauth2:%s' "$(cat /run/secrets/NEMO_CI_TRIAGE_TOKEN)" | base64 -w0)" \
31+
GIT_CONFIG_VALUE_0="Authorization: Basic $(printf 'oauth2:%s' "$(cat /run/secrets/CERNO_TOKEN)" | base64 -w0)" \
3232
uv pip install --no-cache-dir \
33-
"nemo-ci-triage @ git+${CI_SERVER_URL}/dl/nemo/nemo-ci-triage.git@${NEMO_CI_TRIAGE_COMMIT}"
33+
"cerno @ git+${CI_SERVER_URL}/dl/nemo/cerno.git@${CERNO_COMMIT}"

‎docs/api-guide/core/generalized_tensor_parallel.md‎

Lines changed: 123 additions & 6 deletions
Large diffs are not rendered by default.

‎megatron/core/dist_checkpointing/serialization.py‎

Lines changed: 1 addition & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -132,15 +132,7 @@ def load(
132132
# params with a high-precision state dict;
133133
# 2. When using delayed scaling, this loading process writes an extra value into the global
134134
# amax_history buffer of Transformer Engine, which is undesirable.
135-
#
136-
# When the sharded strategy supports per-tensor streaming dequantize
137-
# (``stream_ckpt_dequant``), both concerns are handled inside the
138-
# LoadPlanner on a per-tensor basis, which avoids peaking GPU memory
139-
# with N simultaneous high-precision scratch tensors before the load
140-
# begins. Covers FP8/MXFP8/blockwise-FP8/NVFP4 via the common
141-
# ``QuantizedTensor`` base class.
142-
if not getattr(sharded_strategy, "stream_ckpt_dequant", False):
143-
force_all_tensors_to_non_fp8(sharded_state_dict)
135+
force_all_tensors_to_non_fp8(sharded_state_dict)
144136

145137
sharded_state_dict, nonpersistent_state_dict, sh_ten_factories = load_preprocess(
146138
sharded_state_dict

‎megatron/core/dist_checkpointing/strategies/fully_parallel.py‎

Lines changed: 2 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -184,13 +184,6 @@ def __init__(
184184
self.cached_distribution: Optional[ShardDistribution] = None
185185
self.cached_global_metadata: Optional[Metadata] = None
186186

187-
@property
188-
def stream_ckpt_dequant(self) -> bool:
189-
"""Forward the streaming dequantize flag from the wrapped strategy so that
190-
``serialization.load`` can skip the upfront ``force_all_tensors_to_non_fp8`` pass
191-
when streaming is enabled."""
192-
return getattr(self.base_strategy, "stream_ckpt_dequant", False)
193-
194187
@debug_time("FullyParallelLoadStrategyWrapper.load", logger)
195188
def load(
196189
self,
@@ -245,11 +238,11 @@ def load(
245238
# Step 3: load part of the checkpoint.
246239
# Load only sharded objects first. ShardedTensors will be loaded separately
247240
# so that we can keep track of sharded tensors loaded by this rank
248-
(sharded_tensors, sharded_state_dict, to_load_shards, unloaded_shards) = (
241+
sharded_tensors, sharded_state_dict, to_load_shards, unloaded_shards = (
249242
self._defer_loading_sharded_tensors(sharded_state_dict)
250243
)
251244

252-
(sharded_objects, sharded_state_dict, to_load_objects, unloaded_objects) = (
245+
sharded_objects, sharded_state_dict, to_load_objects, unloaded_objects = (
253246
self._defer_loading_sharded_objects(sharded_state_dict)
254247
)
255248

0 commit comments

Comments
 (0)