radixark · fzyzcjy · Aug 9, 2026 · Jul 28, 2026 · Jul 28, 2026 · Jul 26, 2026
diff --git a/docker/npu_patch/miles.patch b/docker/npu_patch/miles.patch
@@ -340,7 +340,7 @@ diff --git a/miles/utils/external_utils/command_utils.py b/miles/utils/external_
 index d016e01ac..08b4d6eff 100644
 --- a/miles/utils/external_utils/command_utils.py
 +++ b/miles/utils/external_utils/command_utils.py
-@@ -193,6 +193,112 @@ def execute_train(
+@@ -193,6 +193,107 @@ def execute_train(
          )
 
 
@@ -434,18 +434,13 @@ index d016e01ac..08b4d6eff 100644
 +    )
 +
 +    if get_bool_env_var("SLIME_SCRIPT_ENABLE_RAY_SUBMIT", "1"):
-+        cmd_megatron_model_source = (
-+            f'source "{repo_base_dir}/scripts/models/{megatron_model_type}.sh" && '
-+            if megatron_model_type is not None
-+            else ""
-+        )
++        model_args = load_model_args(megatron_model_type) if megatron_model_type is not None else ""
 +        exec_command_cpu(
 +            f"export no_proxy=127.0.0.1 && export PYTHONUNBUFFERED=1 && "
-+            f"{cmd_megatron_model_source}"
 +            f'ray job submit --address="http://127.0.0.1:8265" '
 +            f"--runtime-env-json='{runtime_env_json}' "
 +            f"-- python3 {train_script} "
-+            f"{'${MODEL_ARGS[@]}' if megatron_model_type is not None else ''} "
++            f"{model_args} "
 +            f"{train_args}"
 +        )
 +

diff --git a/docs/advanced/on-policy-distillation.md b/docs/advanced/on-policy-distillation.md
@@ -144,7 +144,8 @@ hf download --repo-type dataset zhuzilin/dapo-math-17k --local-dir /root/dapo-ma
 
 # 2. Convert student model
 cd /root/miles
-source scripts/models/qwen3-8B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-8B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
     ${MODEL_ARGS[@]} \
     --hf-checkpoint /root/Qwen3-8B \

diff --git a/docs/examples/openhermes-sft.md b/docs/examples/openhermes-sft.md
@@ -27,7 +27,8 @@ If you don't already have it:
 hf download Qwen/Qwen3-4B-Base --local-dir /root/Qwen3-4B-Base
 
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3-4B-Base \

diff --git a/docs/examples/reproducibility.md b/docs/examples/reproducibility.md
@@ -70,7 +70,8 @@ hf download --repo-type dataset openai/gsm8k         --local-dir /root/gsm8k
 hf download Qwen/Qwen2.5-0.5B-Instruct               --local-dir /root/Qwen2.5-0.5B-Instruct
 
 cd /root/miles
-source scripts/models/qwen2.5-0.5B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen2.5-0.5B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen2.5-0.5B-Instruct \

diff --git a/docs/examples/search-r1.md b/docs/examples/search-r1.md
@@ -56,7 +56,8 @@ python $WORK_DIR/scripts/data_process/qa_search_train_merge.py \
 ```bash
 hf download Qwen/Qwen2.5-3B --local-dir /root/Qwen2.5-3B
 cd /root/miles
-source scripts/models/qwen2.5-3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen2.5-3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen2.5-3B \

diff --git a/docs/getting-started/quick-start.md b/docs/getting-started/quick-start.md
@@ -65,7 +65,8 @@ map the HuggingFace weights into a sharded `torch_dist` checkpoint.
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \

diff --git a/docs/models/deepseek/deepseek-v4-flash.md b/docs/models/deepseek/deepseek-v4-flash.md
@@ -88,7 +88,8 @@ python tools/fp8_cast_bf16.py \
    --input-fp8-hf-path /root/models/DeepSeek-V4-Flash-FP8 \
    --output-bf16-hf-path /root/models/DeepSeek-V4-Flash-FP8-bf16/
 
-source scripts/models/deepseek-v4-flash.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py deepseek-v4-flash)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun \
    --nproc-per-node 4 --nnodes 8 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \

diff --git a/docs/models/deepseek/deepseek.md b/docs/models/deepseek/deepseek.md
@@ -52,7 +52,8 @@ Then convert BF16 HF → Megatron `torch_dist`. Run on **4 separate nodes** (`NO
 
 ```bash
 cd miles/
-source scripts/models/deepseek-v3.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py deepseek-v3)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM/ torchrun \
    --nproc-per-node 8 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \

diff --git a/docs/models/glm/glm4-5.md b/docs/models/glm/glm4-5.md
@@ -49,7 +49,8 @@ The bash launcher does **not** convert for you — produce `$BASE_DIR/GLM-4.5-35
 
 ```bash
 cd /root/miles
-source scripts/models/glm4.5-355B-A32B.sh
+MODEL_ARGS_LINE="$(python3 scripts/model_args.py glm4.5-355B-A32B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \

diff --git a/docs/models/glm/glm4-7-flash.md b/docs/models/glm/glm4-7-flash.md
@@ -35,7 +35,8 @@ The bash launcher hardcodes `BASE_DIR=/root/shared`. The Python launcher downloa
 
 ```bash
 cd /root/miles
-source scripts/models/glm4.7-flash.sh
+MODEL_ARGS_LINE="$(python3 scripts/model_args.py glm4.7-flash)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \

diff --git a/docs/models/glm/glm4.md b/docs/models/glm/glm4.md
@@ -32,7 +32,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/glm4-9B.sh
+MODEL_ARGS_LINE="$(python3 scripts/model_args.py glm4-9B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/GLM-Z1-9B-0414 \

diff --git a/docs/models/kimi/kimi-k2.5.md b/docs/models/kimi/kimi-k2.5.md
@@ -2,7 +2,7 @@
 title: Kimi K2.5 / K2.6
 description: Launch recipe for Kimi-K2.5, running full-parameter GRPO on 32 × 8 H200 with an INT4 actor and a BF16 reference.
 ---
-The reference launcher is [`scripts/run-kimi-k25.sh`](https://github.com/radixark/miles/blob/main/scripts/run-kimi-k25.sh), which sources the shared model definition in `scripts/models/kimi-k2-thinking.sh`.
+The reference launcher is [`scripts/run-kimi-k25.sh`](https://github.com/radixark/miles/blob/main/scripts/run-kimi-k25.sh), which loads the shared model definition from `scripts/models/kimi-k2-thinking.py`.
 
 ## 1. Model Introduction
 
@@ -79,7 +79,7 @@ ray start --address=${MASTER_ADDR}:6379 --num-gpus 8 --node-ip-address ${WORKER_
 
 ## 4. Script breakdown
 
-The launcher groups its flags into the arrays that are passed to `train.py`. The model shape comes from `MODEL_ARGS`, which is sourced from `scripts/models/kimi-k2-thinking.sh`. That definition sets the MLA latent ranks (`q_lora_rank=1536`, `kv_lora_rank=512`, `qk_head_dim=128`, `qk_pos_emb_head_dim=64`, `v_head_dim=128`), the MoE routing (384 experts, top-8, sigmoid pre-softmax scoring, FP32 router, `--moe-router-topk-scaling-factor 2.827`), and RoPE (`--rotary-base 50000`, `--rotary-scaling-factor 64.0`). The K2.5 recipe then layers the following on top:
+The launcher groups its flags into the arrays that are passed to `train.py`. The model shape comes from `MODEL_ARGS`, which is loaded from `scripts/models/kimi-k2-thinking.py`. That definition sets the MLA latent ranks (`q_lora_rank=1536`, `kv_lora_rank=512`, `qk_head_dim=128`, `qk_pos_emb_head_dim=64`, `v_head_dim=128`), the MoE routing (384 experts, top-8, sigmoid pre-softmax scoring, FP32 router, `--moe-router-topk-scaling-factor 2.827`), and RoPE (`--rotary-base 50000`, `--rotary-scaling-factor 64.0`). The K2.5 recipe then layers the following on top:
 
 - **`CKPT_ARGS`** wires up the dual checkpoint (INT4 actor via `--hf-checkpoint`, BF16 reference via `--ref-load`) together with `--megatron-to-hf-mode bridge` and `--model-name kimi_k25`.
 - **`ROLLOUT_ARGS`** and **`EVAL_ARGS`** configure GRPO sampling and periodic AIME evaluation (covered in §5.2).

diff --git a/docs/models/kimi/kimi-k2.md b/docs/models/kimi/kimi-k2.md
@@ -47,7 +47,8 @@ Convert across 4 nodes (mirror the DeepSeek-V3 procedure):
 
 ```bash
 cd /root/miles
-source scripts/models/kimi-k2.sh   # or kimi-k2-thinking.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py kimi-k2)" || exit 1   # or kimi-k2-thinking
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM/ torchrun \
    --nproc-per-node 8 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \

diff --git a/docs/models/kimi/moonlight.md b/docs/models/kimi/moonlight.md
@@ -33,7 +33,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/moonlight.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py moonlight)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \

diff --git a/docs/models/mimo/mimo.md b/docs/models/mimo/mimo.md
@@ -33,7 +33,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/mimo-7B-rl.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py mimo-7B-rl)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/MiMo-7B-RL \

diff --git a/docs/models/nemotron/nemotron-3-nano-moe.md b/docs/models/nemotron/nemotron-3-nano-moe.md
@@ -125,7 +125,7 @@ memory pressure rises.
 
 ### 5.5 Notable quirks
 
-From `scripts/models/nemotron-3-nano-30b-a3b.sh` and `scripts/run-nemotron-3-nano-30b-a3b.sh`:
+From `scripts/models/nemotron-3-nano-30b-a3b.py` and `scripts/run-nemotron-3-nano-30b-a3b.sh`:
 
 - **No `--spec`**: AutoBridge + the NemotronH shim synthesize the Megatron MoE spec from HF config.
 - 128 experts, `--moe-router-topk 6`, shared expert (3712-dim).

diff --git a/docs/models/nemotron/nemotron-3-nano.md b/docs/models/nemotron/nemotron-3-nano.md
@@ -112,7 +112,7 @@ memory pressure rises.
 
 ### 5.5 Notable quirks
 
-From `scripts/models/nemotron-3-nano-4b.sh` and `scripts/run-nemotron-3-nano-4b.sh`:
+From `scripts/models/nemotron-3-nano-4b.py` and `scripts/run-nemotron-3-nano-4b.sh`:
 
 - **No `--spec`**: the AutoBridge synthesizes the Megatron spec from HF config.
 - `--position-embedding-type none` (no RoPE).

diff --git a/docs/models/qwen/qwen3-5-moe.md b/docs/models/qwen/qwen3-5-moe.md
@@ -32,7 +32,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.5-35B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.5-35B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
@@ -98,7 +99,7 @@ CPU Adam is enabled (`--optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --
 ### 5.5 Notable quirks
 
 - The Megatron side uses `--moe-token-dispatcher-type flex`; DeepEP isn't enabled here, unlike Qwen3-Next.
-- The model config (`scripts/models/qwen3.5-35B-A3B.sh`) reuses the Qwen3.5 spec: `--attention-output-gate`, `--rotary-base 10000000`, `--rotary-percent 0.25`, `A_log` kept in FP32 via the bridge. See [Backends Beyond Megatron](/advanced/architecture-support).
+- The model config (`scripts/models/qwen3.5-35B-A3B.py`) reuses the Qwen3.5 spec: `--attention-output-gate`, `--rotary-base 10000000`, `--rotary-percent 0.25`, `A_log` kept in FP32 via the bridge. See [Backends Beyond Megatron](/advanced/architecture-support).
 
 ## 6. Pairs Well With
 

diff --git a/docs/models/qwen/qwen3-5.md b/docs/models/qwen/qwen3-5.md
@@ -34,7 +34,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.5-4B.sh   # or qwen3.5-9B.sh / qwen3.5-27B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.5-4B)" || exit 1   # or qwen3.5-9B / qwen3.5-27B
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3.5-4B \
@@ -92,7 +93,7 @@ Only the 27 B script enables CPU Adam (`--optimizer-cpu-offload --overlap-cpu-op
 
 ### 5.5 Notable quirks
 
-From `scripts/models/qwen3.5-4B.sh` (and analogous configs for 9 B / 27 B):
+From `scripts/models/qwen3.5-4B.py` (and analogous configs for 9 B / 27 B):
 
 - `--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec` — attention-output gate, `A_log` parameter handling.
 - `--rotary-base 10000000`, `--rotary-percent 0.25`.

diff --git a/docs/models/qwen/qwen3-6-moe.md b/docs/models/qwen/qwen3-6-moe.md
@@ -47,7 +47,8 @@ hf download Qwen/Qwen3.6-35B-A3B --local-dir /root/models/Qwen3.6-35B-A3B
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.6-35B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.6-35B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
@@ -136,7 +137,7 @@ CPU Adam is enabled (`--optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --
 
 ### 5.5 Notable quirks
 
-From `scripts/models/qwen3.6-35B-A3B.sh` and `scripts/run_qwen3_6_35b_a3b_mtp.py`:
+From `scripts/models/qwen3.6-35B-A3B.py` and `scripts/run_qwen3_6_35b_a3b_mtp.py`:
 
 - `--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec` — Qwen3.6 reuses the Qwen3.5 spec.
 - 256 experts, `--moe-router-topk 8`, `--moe-router-score-function softmax`.

diff --git a/docs/models/qwen/qwen3-6.md b/docs/models/qwen/qwen3-6.md
@@ -45,7 +45,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3.6-27B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3.6-27B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3.6-27B \
@@ -109,7 +110,7 @@ CPU Adam is enabled (`--optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --
 
 ### 5.5 Notable quirks
 
-From `scripts/models/qwen3.6-27B.sh`:
+From `scripts/models/qwen3.6-27B.py`:
 
 - `--spec miles_plugins.models.qwen3_5 get_qwen3_5_spec` — Qwen3.6 reuses the Qwen3.5 spec (gated attention, FP32 `A_log`).
 - `--rotary-base 10000000`, `--rotary-percent 0.25`.

diff --git a/docs/models/qwen/qwen3-moe.md b/docs/models/qwen/qwen3-moe.md
@@ -47,7 +47,8 @@ hf download Qwen/Qwen3-235B-A22B-FP8 --local-dir $BASE_FOLDER/Qwen3-235B-A22B-FP
 ### 3.3 HF → Megatron `torch_dist` conversion
 
 ```bash
-source scripts/models/qwen3-30B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-30B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \

diff --git a/docs/models/qwen/qwen3-next.md b/docs/models/qwen/qwen3-next.md
@@ -42,7 +42,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir $BASE_FOLDER/
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-next-80B-A3B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-next-80B-A3B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM torchrun --nproc-per-node 8 \
    tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \

diff --git a/docs/models/qwen/qwen3.md b/docs/models/qwen/qwen3.md
@@ -39,7 +39,8 @@ hf download --repo-type dataset zhuzilin/aime-2024     --local-dir /root/aime-20
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 PYTHONPATH=/root/Megatron-LM python tools/convert_hf_to_torch_dist.py \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint /root/Qwen3-4B \
@@ -57,9 +58,9 @@ cd /root/miles
 bash scripts/run-qwen3-4B.sh
 ```
 
-Other variants follow the same pattern — replace the script name (`run-qwen3-32B.sh`, etc.) and the `qwen3-XB.sh` model config.
+Other variants follow the same pattern — replace the script name (`run-qwen3-32B.sh`, etc.) and the `qwen3-XB.py` model config.
 
-The Qwen3-4B-Instruct-2507 config (`scripts/models/qwen3-4B-Instruct-2507.sh`) just sets `MODEL_ARGS_ROTARY_BASE=5000000` and re-sources `qwen3-4B.sh` — source it when converting / launching the Instruct-2507 checkpoint.
+The Qwen3-4B-Instruct-2507 config (`scripts/models/qwen3-4B-Instruct-2507.py`) just calls `qwen3-4B` with `rotary_base=5000000` (`MODEL_ARGS_ROTARY_BASE` still works as an environment override) — load it when converting / launching the Instruct-2507 checkpoint.
 
 ## 5. Recipe Configuration
 

diff --git a/docs/models/thinkingmachines/inkling-small.md b/docs/models/thinkingmachines/inkling-small.md
@@ -39,7 +39,7 @@ python scripts/run_inkling.py train \
    --sglang-context-length 4096 --rollout-max-response-len 2048
 ```
 
-The model definition lives in `scripts/models/inkling-small.sh` (`MODEL_ARGS_NUM_LAYERS` overrides the layer count for sliced smoke/parity checkpoints). HF → `torch_dist` conversion uses the same tool as Inkling with this recipe file — a single 8-GPU node (TP8 EP8) converts it in one pass.
+The model definition lives in `scripts/models/inkling-small.py` (`MODEL_ARGS_NUM_LAYERS` overrides the layer count for sliced smoke/parity checkpoints). HF → `torch_dist` conversion uses the same tool as Inkling with this recipe file — a single 8-GPU node (TP8 EP8) converts it in one pass.
 
 ## 4. Validated parallelism
 

diff --git a/docs/models/thinkingmachines/inkling.md b/docs/models/thinkingmachines/inkling.md
@@ -72,11 +72,12 @@ Pass `--hf-checkpoint <path>` to the launcher when the weights are already on a
 
 ### 4.2 HF → Megatron `torch_dist` conversion
 
-Inkling ships in BF16, so conversion is a single distributed `torch_dist` shard (no precision cast). The model definition comes from `scripts/models/inkling.sh`:
+Inkling ships in BF16, so conversion is a single distributed `torch_dist` shard (no precision cast). The model definition comes from `scripts/models/inkling.py`:
 
 ```bash
 cd /root/miles
-source scripts/models/inkling.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py inkling)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 CONVERT_KEEP_PP1=1 PYTHONPATH=/root/Megatron-LM torchrun \
    --nproc-per-node 4 --nnodes 4 \
    --master-addr ${MASTER_ADDR} --master-port 12345 \

diff --git a/docs/platforms/amd.md b/docs/platforms/amd.md
@@ -68,7 +68,8 @@ ROCm converter is in development.
 
 ```bash
 cd /root/miles
-source scripts/models/qwen3-4B.sh
+MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py qwen3-4B)" || exit 1
+read -ra MODEL_ARGS <<< "${MODEL_ARGS_LINE}"
 MEGATRON_LM_PATH=$(pip list | grep megatron-core | awk '{print $NF}')
 
 PYTHONPATH=${MEGATRON_LM_PATH} python tools/convert_hf_to_torch_dist.py \

diff --git a/docs/user-guide/argument-groups.md b/docs/user-guide/argument-groups.md
@@ -11,7 +11,7 @@ when you need the full default and type for an individual flag.
 
 | Group | Owns | Typical source |
 |---|---|---|
-| [`MODEL_ARGS`](#model-args) | Architecture constants and plugin specs | `scripts/models/<family>.sh` |
+| [`MODEL_ARGS`](#model-args) | Architecture constants and plugin specs | `scripts/models/<family>.py` |
 | [`CKPT_ARGS`](#ckpt-args) | Actor, reference, HF tokenizer/config, save paths | Launch script |
 | [`ROLLOUT_ARGS`](#rollout-args) | Prompt data, sampling, reward, train/eval batch flow | Launch script |
 | [`EVAL_ARGS`](#eval-args) | Evaluation datasets and eval-only sampling overrides | Launch script |
@@ -24,7 +24,7 @@ when you need the full default and type for an individual flag.
 ## MODEL_ARGS - architecture constants
 
 `MODEL_ARGS` tells Megatron what model it is instantiating. Megatron cannot infer all
-architecture details from a HuggingFace checkpoint, so each recipe sources a matching
+architecture details from a HuggingFace checkpoint, so each recipe loads a matching
 file from `scripts/models/`.
 
 Common entries: