Cl 834 #5145
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -1,2 +1,4 @@ | ||
| .git | ||
| maxtext_venv | ||
| .venv | ||
| venv13 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change | ||||
|---|---|---|---|---|---|---|
| @@ -0,0 +1,209 @@ | ||||||
| #!/bin/bash | ||||||
| set -e | ||||||
|
|
||||||
| # Activate Python virtual environment | ||||||
| source /usr/local/google/home/muskansh/maxtext_env/bin/activate | ||||||
|
|
||||||
| # --- Environment Variables --- | ||||||
| export PROJECT_ID="tpu-prod-env-one-vm" | ||||||
| export CLUSTER_NAME="v6e-256-c2b3-b478935789" | ||||||
| export ZONE="us-central2" | ||||||
|
|
||||||
| # --- Configuration & Automated Image Build --- | ||||||
| TIMESTAMP=$(date +%m%d%H%M%S) | ||||||
| export WORKLOAD_IMAGE="gcr.io/tpu-prod-env-one-vm/param3_21jul:muskansh_${TIMESTAMP}" | ||||||
| export WORKLOAD_NAME="muskansh-qn80b-${TIMESTAMP}" | ||||||
| export DEVICE_TYPE="v6e-256" | ||||||
| export NUM_SLICES=1 | ||||||
| export PRIORITY="very-high" | ||||||
| export MAX_RESTARTS=0 | ||||||
| export NUM_STEPS=15 | ||||||
| export MODEL_NAME="qwen3-next-80b-a3b" | ||||||
| export BASE_OUTPUT_DIR="/tmp/qwen3-next-80b-profiles/run-${TIMESTAMP}" | ||||||
|
|
||||||
| echo "========================================================================" | ||||||
| echo "Building and uploading Docker runner image from /usr/local/google/home/muskansh/maxtext" | ||||||
| echo "Target Image: ${WORKLOAD_IMAGE}" | ||||||
| echo "========================================================================" | ||||||
|
|
||||||
| ( | ||||||
| cd /usr/local/google/home/muskansh/maxtext && \ | ||||||
| CLOUD_IMAGE_NAME="${WORKLOAD_IMAGE}" \ | ||||||
| BASE_IMAGE="gcr.io/tpu-prod-env-one-vm/param3_21jul:latest" \ | ||||||
| bash src/dependencies/scripts/docker_upload_runner.sh | ||||||
| ) | ||||||
|
|
||||||
| echo "Docker image upload complete: ${WORKLOAD_IMAGE}" | ||||||
|
|
||||||
| # --- XLA Flags --- | ||||||
| XLA_FLAGS_ARRAY=( | ||||||
| "--xla_msa_enable_sync_slice_replacement=false" | ||||||
| "--xla_tpu_enable_sparse_core_collective_offload_2d_all_gather=true" | ||||||
| "--xla_tpu_enable_sparse_core_collective_offload_reduce_scatter=true" | ||||||
| "--xla_msa_enable_sync_copy_replacement=false" | ||||||
| "--xla_tpu_scoped_vmem_limit_kib=81000" | ||||||
| "--xla_tpu_enable_sparse_core_collective_offload_all_gather=true" | ||||||
| "--xla_tpu_enable_sparse_core_collective_offload_all_reduce=true" | ||||||
| "--xla_tpu_enable_concurrent_sparse_core_offloading=true" | ||||||
| "--xla_tpu_enable_sparse_core_offload_queuing_in_lhs=true" | ||||||
| "--xla_tpu_enable_layer_scheduler_for_dependent_collectives=true" | ||||||
| "--xla_tpu_use_single_sparse_core_for_all_gather_offload=true" | ||||||
| "--xla_tpu_sparse_core_all_gather_latency_multiplier=1" | ||||||
| "--xla_tpu_sparse_core_reduce_scatter_latency_multiplier=3" | ||||||
| "--xla_tpu_offload_gather_to_sparsecore=true" | ||||||
| "--xla_tpu_dvfs_p_state=7" | ||||||
| "--xla_tpu_disable_sparse_core_collective_offload_remover=true" | ||||||
| "--xla_tpu_use_tc_device_shape_on_sc=true" | ||||||
| "--xla_sc_enable_instruction_fusion=false" | ||||||
| "--xla_sc_disable_megacore_partitioning=true" | ||||||
| "--xla_tpu_enable_async_collective_fusion=true" | ||||||
| "--xla_tpu_overlap_compute_collective_tc=true" | ||||||
| "--xla_tpu_enable_async_collective_fusion_multiple_steps=true" | ||||||
| "--xla_tpu_enable_async_collective_fusion_fuse_all_gather=false" | ||||||
| "--xla_tpu_enable_async_collective_fusion_fuse_reduce_scatter=false" | ||||||
| "--xla_tpu_enable_async_collective_fusion_fuse_all_reduce=false" | ||||||
| "--xla_tpu_enable_latency_hiding_scheduler=true" | ||||||
| "--xla_latency_hiding_scheduler_rerun=10" | ||||||
| "--xla_tpu_all_gather_collective_matmul_mode=post_spmd_conservative" | ||||||
| "--xla_tpu_reduce_scatter_collective_matmul_mode=post_spmd_conservative" | ||||||
| "--xla_latency_hiding_scheduler_enable_selective_resources=true" | ||||||
| "--xla_tpu_enable_scheduler_memory_pressure_tracking=true" | ||||||
| "--xla_tpu_host_transfer_overlap_limit=4" | ||||||
| "--xla_tpu_aggressive_opt_barrier_removal=ENABLED" | ||||||
| "--xla_lhs_prioritize_async_depth_over_stall=ENABLED" | ||||||
| "--xla_should_allow_loop_variant_parameter_in_chain=ENABLED" | ||||||
| "--xla_should_add_loop_invariant_op_in_chain=ENABLED" | ||||||
| "--xla_max_concurrent_host_send_recv=100" | ||||||
| "--xla_tpu_rerun_latency_hiding_scheduler_post_sc_assignment=true" | ||||||
| "--xla_tpu_scheduler_percent_shared_memory_limit=150" | ||||||
| ) | ||||||
| export XLA_FLAGS="${XLA_FLAGS_ARRAY[*]}" | ||||||
|
|
||||||
| # --- MaxText Workload Overrides --- | ||||||
| MAXTEXT_ARGS_ARRAY=( | ||||||
| "model_name=${MODEL_NAME}" | ||||||
| "base_output_directory=${BASE_OUTPUT_DIR}" | ||||||
| "run_name=${WORKLOAD_NAME}" | ||||||
| "dataset_type=synthetic" | ||||||
| "dataset_name=synthetic" | ||||||
| "dtype=bfloat16" | ||||||
| "allow_split_physical_axes=True" | ||||||
| "ici_expert_parallelism=4" | ||||||
| "use_ring_of_experts=True" | ||||||
| "custom_mesh=hybrid_ring_64x4" | ||||||
| "use_ragged_sort=True" | ||||||
| "use_random_routing=False" | ||||||
| "per_device_batch_size=4" | ||||||
| "num_moe_token_chunks=2" | ||||||
| "opt_type=muon" | ||||||
| "max_target_length=2048" | ||||||
| "ragged_buffer_factor=1.5" | ||||||
| "remat_policy=custom" | ||||||
| "decoder_layer_input=device" | ||||||
| "context=device" | ||||||
| "reuse_example_batch=1" | ||||||
| "ici_fsdp_parallelism=-1" | ||||||
| "steps=${NUM_STEPS}" | ||||||
| "sa_block_q=1024" | ||||||
| "sa_block_kv=1024" | ||||||
| "sa_block_kv_compute=512" | ||||||
| "sa_block_q_dkv=1024" | ||||||
| "sa_block_kv_dkv=1024" | ||||||
| "sa_block_kv_dkv_compute=1024" | ||||||
| "sa_fuse_reciprocal=false" | ||||||
| "use_splash_scheduler=true" | ||||||
| "sa_use_base2_exp=true" | ||||||
| "dq_reduction_steps=3" | ||||||
| "hardware=tpu" | ||||||
| "skip_jax_distributed_system=False" | ||||||
| "attention=flash" | ||||||
| "use_tokamax_splash=True" | ||||||
| "sa_use_fused_bwd_kernel=True" | ||||||
| "sparse_matmul=True" | ||||||
| "megablox=True" | ||||||
| "wi_tile_fwd_batch_seq=512" | ||||||
| "wi_tile_dlhs_batch_seq=512" | ||||||
| "wi_tile_drhs_batch_seq=512" | ||||||
| "wo_tile_fwd_batch_seq=512" | ||||||
| "wo_tile_dlhs_batch_seq=512" | ||||||
| "wo_tile_drhs_batch_seq=512" | ||||||
| "wi_tile_fwd_embed_dim=3072" | ||||||
| "wi_tile_fwd_mlp_dim=1536" | ||||||
| "wi_tile_dlhs_embed_dim=3072" | ||||||
| "wi_tile_dlhs_mlp_dim=1536" | ||||||
| "wi_tile_drhs_embed_dim=3072" | ||||||
| "wi_tile_drhs_mlp_dim=1536" | ||||||
| "wo_tile_fwd_embed_dim=3072" | ||||||
| "wo_tile_fwd_mlp_dim=1536" | ||||||
| "wo_tile_dlhs_embed_dim=3072" | ||||||
| "wo_tile_dlhs_mlp_dim=1536" | ||||||
| "wo_tile_drhs_embed_dim=3072" | ||||||
| "wo_tile_drhs_mlp_dim=1536" | ||||||
| "use_tokamax_gmm=True" | ||||||
| "use_gmm_v2=True" | ||||||
| "optimizer_memory_host_offload=False" | ||||||
| "parameter_memory_host_offload=False" | ||||||
| "enable_checkpointing=False" | ||||||
| "async_checkpointing=False" | ||||||
| "tokenizer_type=huggingface" | ||||||
| "tokenizer_path=assets/tokenizers/qwen3-tokenizer" | ||||||
| "override_model_config=true" | ||||||
| "mhc_expansion_rate=4" | ||||||
| "enable_mhc_lite=True" | ||||||
| "use_mhc_pallas_kernel=True" | ||||||
| "mhc_pallas_kernel_fwd_block_size=256" | ||||||
| "mhc_pallas_kernel_bwd_block_size=256" | ||||||
| "use_gdn_kernel=True" | ||||||
| "use_hybrid_gdn=True" | ||||||
| "profiler=xplane" | ||||||
| "profiler_steps=2" | ||||||
| "skip_first_n_steps_for_profiler=1" | ||||||
| "enable_tpu_profiling_options=True" | ||||||
| "upload_all_profiler_results=False" | ||||||
| "enable_tensorboard=False" | ||||||
| "abort_on_nan_loss=False" | ||||||
| "abort_on_inf_loss=False" | ||||||
| ) | ||||||
| MAXTEXT_ARGS="${MAXTEXT_ARGS_ARRAY[*]}" | ||||||
|
|
||||||
| USER_TOKEN=$(gcloud auth application-default print-access-token 2>/dev/null || gcloud auth print-access-token 2>/dev/null || true) | ||||||
|
|
||||||
| # The command to run inside the container | ||||||
| RUN_COMMAND="set -e && \ | ||||||
| export LIBTPU_INIT_ARGS=\"${XLA_FLAGS}\" && \ | ||||||
| export JAX_PLATFORMS='tpu,cpu' && \ | ||||||
| export ENABLE_PJRT_COMPATIBILITY='true' && \ | ||||||
| export JAX_DISTRIBUTED_INITIALIZE_TIMEOUT=1800 && \ | ||||||
| export PYTHONPATH=/deps:/deps/src:/deps/src/maxtext/src && \ | ||||||
| export CLOUDSDK_AUTH_ACCESS_TOKEN='${USER_TOKEN}' && \ | ||||||
| python3 src/maxtext/trainers/pre_train/train.py src/maxtext/configs/base.yml ${MAXTEXT_ARGS} && \ | ||||||
| (python3 -c 'import os, glob; from google.cloud import storage; import google.oauth2.credentials; token = os.environ.get(\"CLOUDSDK_AUTH_ACCESS_TOKEN\"); client = storage.Client(credentials=google.oauth2.credentials.Credentials(token), project=\"tpu-prod-env-one-vm\") if token else storage.Client(project=\"tpu-prod-env-one-vm\"); bucket = client.bucket(\"muskansh-data\"); [bucket.blob(f\"qwen3-next-80b-profiles/{os.path.relpath(p, \"/tmp/qwen3-next-80b-profiles\")}\").upload_from_filename(p) for p in glob.glob(\"/tmp/qwen3-next-80b-profiles/**/*\", recursive=True) if os.path.isfile(p)]' || true) && \ | ||||||
| sleep 3600" | ||||||
|
There was a problem hiding this comment. The It would be better to make this behavior optional or configurable. For example, you could add an environment variable
Suggested change
|
||||||
|
|
||||||
| # --- XPK Workload Creation --- | ||||||
| echo "Creating XPK workload: ${WORKLOAD_NAME} on cluster: ${CLUSTER_NAME}" | ||||||
|
|
||||||
| PYTHONPATH=/usr/local/google/home/muskansh/xpk/src python3 -P -m xpk.main workload create \ | ||||||
| --cluster="${CLUSTER_NAME}" \ | ||||||
| --project="${PROJECT_ID}" \ | ||||||
| --zone="${ZONE}" \ | ||||||
| --priority="${PRIORITY}" \ | ||||||
| --max-restarts="${MAX_RESTARTS}" \ | ||||||
| --device-type="${DEVICE_TYPE}" \ | ||||||
| --num-slices="${NUM_SLICES}" \ | ||||||
| --docker-image="${WORKLOAD_IMAGE}" \ | ||||||
| --enable-debug-logs \ | ||||||
| --workload="${WORKLOAD_NAME}" \ | ||||||
| --command="${RUN_COMMAND}" | ||||||
|
|
||||||
| LOGS_URL="https://console.cloud.google.com/logs/query;query=resource.type%3D%22k8s_container%22%0Aresource.labels.project_id%3D%22${PROJECT_ID}%22%0Aresource.labels.location%3D%22${ZONE}%22%0Aresource.labels.cluster_name%3D%22${CLUSTER_NAME}%22%0Aresource.labels.namespace_name%3D%22default%22%0Aresource.labels.pod_name%3A%22${WORKLOAD_NAME}-slice-job-0-0-%22%0Aseverity%3E%3DDEFAULT;storageScope=project;duration=P1D?project=${PROJECT_ID}" | ||||||
| GKE_URL="https://console.cloud.google.com/kubernetes/service/${ZONE}/${CLUSTER_NAME}/default/${WORKLOAD_NAME}/details?project=${PROJECT_ID}" | ||||||
|
|
||||||
| echo "========================================================================" | ||||||
| echo "📋 Pantheon Cloud Logging (Worker 0 Logs):" | ||||||
| echo "${LOGS_URL}" | ||||||
| echo "" | ||||||
| echo "☸️ GKE Workload Details:" | ||||||
| echo "${GKE_URL}" | ||||||
| echo "" | ||||||
| echo "========================================================================" | ||||||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,87 @@ | ||
| # Copyright 2025 Google LLC | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # https://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| # maxtext/configs/models/qwen3-next-80b-a3b-256e.yml | ||
|
|
||
| # Set the decoder block to our new implementation | ||
| decoder_block: "qwen3_next" | ||
|
|
||
| # Core Architectural Parameters | ||
| base_emb_dim: 3072 | ||
| base_num_decoder_layers: 40 | ||
| base_num_query_heads: 64 | ||
| base_num_kv_heads: 8 | ||
| head_dim: 64 | ||
| vocab_size: 128008 | ||
| normalization_layer_epsilon: 1.0e-6 | ||
|
|
||
| # MoE Specific Parameters | ||
| # base_mlp_dim sizes the dense-prefix layer's MLP (see first_num_dense_layers below); | ||
| # base_moe_mlp_dim sizes every other (MoE) layer's routed + shared experts. | ||
| base_mlp_dim: 10240 | ||
| base_moe_mlp_dim: 1536 | ||
| num_experts: 256 | ||
| shared_experts: 1 | ||
| num_experts_per_tok: 8 | ||
| norm_topk_prob: true | ||
| # Router parity with reference: fp32 gate-logit matmul. | ||
| float32_gate_logits: true | ||
|
|
||
| # DeepSeek-V3-style router: sigmoid scoring, aux-loss-free expert-bias balancing | ||
| # (primary mechanism) plus a small complementary aux loss (secondary safety net, | ||
| # matching DeepSeek-V3's own combined approach), and a top-k weight scaling factor. | ||
| routed_score_func: "sigmoid" | ||
| routed_bias: true | ||
| routed_bias_update_rate: 1.0e-3 | ||
| routed_scaling_factor: 2.5 | ||
| load_balance_loss_weight: 1.0e-3 | ||
|
|
||
| # Explicit (diverges from the reference's False): keeps the shared-expert gate | ||
| # that real Qwen3-Next uses. | ||
| moe_shared_expert_gate: true | ||
|
|
||
| # The first layer is a dense MLP (no MoE) and always uses full attention, | ||
| # mirroring DeepSeek V3's dense-prefix pattern. | ||
| first_num_dense_layers: 1 | ||
|
|
||
| # Qwen3-Next Specific Parameters for Linear Attention (Gated Delta Net). | ||
| # Attention schedule (2 GDN : 1 full-attention, full-attention first in each cycle): | ||
| # [0,1,1,0,1,1,0,1,1,...] where 0=full attention, 1=Gated Delta Net. | ||
| inhomogeneous_layer_cycle_interval: 3 | ||
| full_attention_layer_offset: 0 | ||
| gdn_conv_kernel_dim: 4 | ||
| gdn_key_head_dim: 128 | ||
| gdn_value_head_dim: 128 | ||
| gdn_num_key_heads: 16 | ||
| gdn_num_value_heads: 32 | ||
| gdn_chunk_size: 64 | ||
| # L2-norm on Q/K inside the Gated Delta Rule is standard practice for linear-attention | ||
| # architectures (stabilizes the recurrent state, since linear attention lacks softmax's | ||
| # implicit boundedness) and matches real Qwen3-Next's own Gated Delta Net. This is also | ||
| # MaxText's own default (base.yml); kept explicit here for clarity. | ||
| use_qk_norm_in_gdn: true | ||
|
|
||
| # RoPE Settings | ||
| rope_max_timescale: 10000 | ||
| partial_rotary_factor: 1 | ||
|
|
||
| # Hyper-connections: mHC-lite enabled with Pallas kernel | ||
| mhc_expansion_rate: 4 | ||
| enable_mhc_lite: true | ||
| use_mhc_pallas_kernel: true | ||
| mhc_pallas_kernel_fwd_block_size: 256 | ||
| mhc_pallas_kernel_bwd_block_size: 256 | ||
|
|
||
| # General Model Settings | ||
| enable_dropout: false |
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This script contains several hardcoded user-specific paths and resource names, which makes it non-portable and difficult for other developers to use.
For example:
source /usr/local/google/home/muskansh/maxtext_env/bin/activatecd /usr/local/google/home/muskansh/maxtextclient.bucket("muskansh-data")PYTHONPATH=/usr/local/google/home/muskansh/xpk/srcThese should be parameterized, for instance, by using environment variables defined at the top of the script or passed as arguments. This would greatly improve the script's reusability.
Suggestion:
At the top of the script, define variables for these paths and names, allowing them to be overridden by environment variables if they are already set:
Then use these variables throughout the script, for example: