Merge pull request #8 from Letian88/devel-1.3.0

feat: default GPU workflows to NVIDIA L20

Author: Zikkying

README.md

@@ -241,7 +241,7 @@ Create `global_bohrium.json` to submit workflows to the Bohrium cloud platform:
 {
   "lammps_image_name": "registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2",
   "lammps_run_command":"lmp -in in.lammps",
-  "scass_type":"c8_m32_1 * NVIDIA 4090"
+  "scass_type":"c16_m120_1 * NVIDIA L20"
 }
 ```
 

apex/account.py

@@ -27,7 +27,7 @@
     "batch_type": "OpenAPI",
     "context_type": "OpenAPI",
     "platform": "ali",
-    "machine_type": "c8_m32_1 * NVIDIA 4090",
+    "machine_type": "c16_m120_1 * NVIDIA L20",
     "output_log": False,
     "dispatcher_image": SANDBOX_DISPATCHER_IMAGE,
     "apex_image_name": "registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post",

apex/skills/apex-flow/SKILL.md

@@ -140,9 +140,9 @@ Options to offer via AskQuestion:
    python scripts/validate_apex_combo.py list-combos --backend lammps --prefer gpu
    python scripts/validate_apex_combo.py check \
      --image registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2 \
-     --scass "c8_m32_1 * NVIDIA 4090"
+     --scass "c16_m120_1 * NVIDIA L20"
   ```
-   Do **not** hardcode an unverified `scass_type`. Prefer `recommend` / `list-combos` output. Known failures include `deepmd-kit:3.1.0`, `3.1.1-cuda12.1`, `3.1.2`, the combination `deepmd-kit:3.1.1` × `NVIDIA T4`, `c4_m16_cpu`, and `c12_m46_1 * NVIDIA T4`. GPU LAMMPS potentials (`deepmd`, `mace`, `nep`) use `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` with RTX 4090. CPU LAMMPS potentials use `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post`; never pair 0.0.2 with a `*_cpu` machine because sequential CPU validation stalled before the container command started. `apex submit` only enforces 0.0.2 for GPU-potential phonon and Grüneisen workflows. Do not emit `plugin load libdeepmd_lmp.so` for 0.0.2.
+   Do **not** hardcode an unverified `scass_type`. Prefer `recommend` / `list-combos` output. Known failures include `deepmd-kit:3.1.0`, `3.1.1-cuda12.1`, `3.1.2`, the combination `deepmd-kit:3.1.1` × `NVIDIA T4`, `c4_m16_cpu`, and `c12_m46_1 * NVIDIA T4`. GPU LAMMPS potentials (`deepmd`, `mace`, `nep`) use `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` with NVIDIA L20 by default; RTX 4090 remains a validated compatible option. CPU LAMMPS potentials use `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post`; never pair 0.0.2 with a `*_cpu` machine because sequential CPU validation stalled before the container command started. `apex submit` only enforces 0.0.2 for GPU-potential phonon and Grüneisen workflows. Do not emit `plugin load libdeepmd_lmp.so` for 0.0.2.
 10. **MUST use the bundled frozen DPA model under** `models/` **for LAMMPS + DeePMD unless the user explicitly requests another compatible model.** The skill ships
    `models/DPA-3.2-5M/DPA-3.2-5M-OMat24.pth`, a ready-to-run frozen
    DPA-3.2-5M OMat24 model. Copy it into the job directory before generating
@@ -340,7 +340,7 @@ See `reference/submission.md` for the full validated template.
 1. **Finite-temperature backend limits**: `finite_t_elastic` and `melting_point` are LAMMPS-only. `finite_t_latt` and `annealing` support LAMMPS and VASP, but not ABACUS. VASP uses `MDALGO=3` and requires a binary compiled with `-Dtbdyn`; annealing `protocol="coexistence"` is a fixed-temperature equilibration plus production run.
 2. **Model files must be in job directory.** For MLIP workflows, the model file (`.pb`, `.pth`, `.model`, etc.) must be present in the submitted directory. Use relative paths in `param.json`. For DeePMD/DPA, copy `models/DPA-3.2-5M/DPA-3.2-5M-OMat24.pth`. Default to `"type_map": "auto"` for every LAMMPS interaction; specify a dictionary only when the user explicitly needs a fixed custom ordering.
 3. **Joint workflow recommended.** Use `joint` flow (relaxation + properties) for most use cases to ensure proper relaxation before property calculations.
-4. **GPU for ML potentials.** DeePMD, MACE, and NEP benefit from GPU acceleration. Set `scass_type` to a validated GPU SKU from `validate_apex_combo.py recommend --prefer gpu` (default: `"c8_m32_1 * NVIDIA 4090"`).
+4. **GPU for ML potentials.** DeePMD, MACE, and NEP benefit from GPU acceleration. Set `scass_type` to a validated GPU SKU from `validate_apex_combo.py recommend --prefer gpu` (default: `"c16_m120_1 * NVIDIA L20"`; RTX 4090 remains compatible).
 5. **Supercell sizing depends on the input atom count, not only the default JSON.**
    Treat defaults as targets for **unit-cell inputs**. First inspect the user's
    structure; if it is already large enough, prefer `[1,1,1]` after confirmation.

apex/skills/apex-flow/data/default_templates.json

@@ -4,8 +4,8 @@
     "backends": {
         "lammps_gpu": {
             "image": "registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2",
-            "machine": "c8_m32_1 * NVIDIA 4090",
-            "machine_sandbox": "c8_m32_1 * NVIDIA 4090",
+            "machine": "c16_m120_1 * NVIDIA L20",
+            "machine_sandbox": "c16_m120_1 * NVIDIA L20",
             "potentials": ["deepmd", "mace", "nep"]
         },
         "lammps_cpu": {
@@ -30,7 +30,7 @@
     },
     "sandbox_machines": {
         "cpu": ["c2_m4_cpu", "c2_m8_cpu", "c8_m32_cpu", "c32_m128_cpu", "c64_m256_cpu"],
-        "gpu": ["c8_m32_1 * NVIDIA 4090", "c16_m64_1 * NVIDIA 4090", "c16_m64_1 * NVIDIA 5090"]
+        "gpu": ["c16_m120_1 * NVIDIA L20", "c8_m32_1 * NVIDIA 4090", "c16_m64_1 * NVIDIA 4090", "c16_m64_1 * NVIDIA 5090"]
     },
     "properties": {
         "eos": {"lammps": true, "abacus": true, "vasp": true},

apex/skills/apex-flow/data/global_bohrium_direct.json

@@ -6,7 +6,7 @@
     "apex_image_name": "registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post",
     "lammps_image_name": "registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2",
     "lammps_run_command": "lmp -in in.lammps",
-    "scass_type": "c8_m32_1 * NVIDIA 4090",
+    "scass_type": "c16_m120_1 * NVIDIA L20",
     "group_size": 1,
     "pool_size": 1
 }

apex/skills/apex-flow/reference/calculators.md

@@ -112,13 +112,15 @@ CPU image (`gap`, `snap`, `rann`, `eam_alloy`, `eam_fs`, `meam`,
 This image uses integrated USER-DEEPMD and CUDA 12.8. Do not add
 `plugin load libdeepmd_lmp.so`; the plugin command is unsupported. Its unified
 `lmp` entry dispatches RTX 4090/L20-class GPUs to the sm89 build and RTX 5090
-to sm120. The 4090 path is validated; test sm120 on a real 5090 before relying
-on that path.
+to sm120. The L20 and RTX 4090 sm89 paths are validated; L20 is the default
+because it has greater resource availability. Test sm120 on a real 5090 before
+relying on that path.
 
 > ⚠️ **Do NOT use `deepmd-kit:3.1.1` with GPU/T4** — known startup / triclinic issues. Prefer `3.1.3` or later.
 
 > **LAMMPS phonon and Grüneisen**: for GPU potentials, `apex submit` forces
-> the DPA4 image above. It is validated on RTX 4090 and includes phonoLAMMPS.
+> the DPA4 image above. It is validated on NVIDIA L20 and RTX 4090 and includes
+> phonoLAMMPS. L20 is the default resource.
 > CPU potentials keep the CPU image. Do not use the DPA4 image on a CPU
 > machine; sequential CPU jobs stalled during container preparation.
 
@@ -650,9 +652,9 @@ Before submit:
 
 | Workload | Bohrium Machine | Rationale |
 |----------|----------------|-----------|
-| LAMMPS + GPU potential (DeePMD/MACE/NEP) | `c8_m32_1 * NVIDIA 4090` | GPU acceleration |
+| LAMMPS + GPU potential (DeePMD/MACE/NEP) | `c16_m120_1 * NVIDIA L20` | Validated sm89 runtime; greater resource availability |
 | LAMMPS + CPU potential (EAM/MEAM/SNAP) | `c16_m32_cpu` | CPU sufficient |
 | ABACUS DFT (small cell <50 atoms) | `c16_m32_cpu` | 8 MPI ranks |
 | ABACUS DFT (large cell 50-200 atoms) | `c32_m128_cpu` | 16-32 MPI ranks |
 | VASP DFT | User choice | Depends on system size |
-| Finite-T MD (long runs) | `c8_m32_1 * NVIDIA 4090` | Long MD = GPU beneficial |
+| Finite-T MD (long runs) | `c16_m120_1 * NVIDIA L20` | Long MD = GPU beneficial; validated sm89 runtime |

apex/skills/apex-flow/reference/lammps_potentials.md

@@ -218,9 +218,9 @@ manual map.
 
 | Potential Type | Recommended Image | GPU? |
 |---------------|-------------------|------|
-| deepmd | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | Yes (RTX 4090) |
-| mace | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | Yes (RTX 4090) |
-| nep | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | Yes (RTX 4090) |
+| deepmd | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | Yes (NVIDIA L20 default; RTX 4090 compatible) |
+| mace | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | Yes (NVIDIA L20 default; RTX 4090 compatible) |
+| nep | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | Yes (NVIDIA L20 default; RTX 4090 compatible) |
 | gap | `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post` | No (CPU) |
 | snap | `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post` | No |
 | rann | `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post` | No |

apex/skills/apex-flow/reference/submission.md

@@ -119,7 +119,7 @@ dflow validates workflow names against RFC 1123 subdomain regex. Names like `"Cu
 | Role | Image | Notes |
 |------|-------|-------|
 | **Outer job (submission client)** | `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post` | Lightweight; just runs `apex submit` |
-| **LAMMPS calculator (GPU potentials)** | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | RTX 4090; includes phonoLAMMPS |
+| **LAMMPS calculator (GPU potentials)** | `registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2` | NVIDIA L20 default; RTX 4090 compatible; includes phonoLAMMPS |
 | **LAMMPS calculator (CPU potentials)** | `registry.dp.tech/dptech/dp/native/prod-397637/apex-flow:1.3.0.post` | EAM/MEAM/SNAP/GAP/RANN CPU backend |
 | **ABACUS calculator** | (same APEX image has ABACUS) | Or user-specified |
 | **VASP calculator** | User must provide after confirming license | Commercial; **never invent a default image** |
@@ -135,7 +135,7 @@ dflow validates workflow names against RFC 1123 subdomain regex. Names like `"Cu
 
 | Backend | scass_type (inner containers) | Notes |
 |---------|-------------------------------|-------|
-| LAMMPS (DeePMD/MACE/NEP) | `c8_m32_1 * NVIDIA 4090` | GPU beneficial |
+| LAMMPS (DeePMD/MACE/NEP) | `c16_m120_1 * NVIDIA L20` | Validated sm89 default; RTX 4090 remains compatible |
 | LAMMPS (EAM/MEAM/SNAP) | `c16_m32_cpu` | CPU sufficient |
 | ABACUS | `c16_m32_cpu` | CPU |
 | VASP | `c32_m128_cpu` (default) | Align `mpirun -n <N>` with CPU count |
@@ -216,7 +216,7 @@ Do not upload or submit unless validation reports `Validation PASSED` and both
 project ID lines report `type=int`. Submit the newly validated directory as a
 new outer Bohrium job; retrying an old outer job reuses its old input snapshot.
 
-> For GPU potentials (DeePMD, MACE, NEP), change `scass_type` to `"c8_m32_1 * NVIDIA 4090"`.
+> For GPU potentials (DeePMD, MACE, NEP), change `scass_type` to `"c16_m120_1 * NVIDIA L20"` by default; RTX 4090 remains compatible.
 > Before submitting, run `scripts/validate_apex_combo.py check` on the chosen image × scass_type.
 
 ## Agent-Managed Submission Workflow (Complete Lifecycle)

apex/skills/apex-flow/scripts/generate_config.py

@@ -217,15 +217,15 @@ def select_lammps_image(potential: str = None) -> str:
 
 # scass_type defaults for inner dflow containers (legacy Bohrium)
 SCASS_TYPES = {
-    "lammps_gpu": "c8_m32_1 * NVIDIA 4090",
+    "lammps_gpu": "c16_m120_1 * NVIDIA L20",
     "lammps_cpu": "c16_m32_cpu",
     "abacus": "c16_m32_cpu",
     "vasp": "c32_m128_cpu",
 }
 
 # machine_type defaults for OpenAPI Sandbox
 SANDBOX_MACHINE_TYPES = {
-    "lammps_gpu": "c8_m32_1 * NVIDIA 4090",
+    "lammps_gpu": "c16_m120_1 * NVIDIA L20",
     "lammps_cpu": "c8_m32_cpu",
     "abacus": "c8_m32_cpu",
     "vasp": "c32_m128_cpu",
@@ -1618,7 +1618,7 @@ def main():
     create.add_argument("--sandbox", action="store_true",
                         help="Use OpenAPI Sandbox mode (access_key auth, no ticket)")
     create.add_argument("--machine-type",
-                        help="Override machine_type for sandbox (e.g. 'c8_m32_1 * NVIDIA 4090')")
+                        help="Override machine_type for sandbox (e.g. 'c16_m120_1 * NVIDIA L20')")
     create.add_argument("--scass-type",
                         help="Override scass_type for inner dflow containers (legacy Bohrium)")
     create.add_argument("--run-command",

apex/skills/apex-flow/scripts/validate_apex_combo.py

@@ -9,7 +9,7 @@
     python validate_apex_combo.py list-combos --backend lammps --prefer gpu
     python validate_apex_combo.py check \\
         --image registry.dp.tech/dptech/dp/native/prod-16664/dpa4-phonolammps:0.0.2 \\
-        --scass "c8_m32_1 * NVIDIA 4090"
+        --scass "c16_m120_1 * NVIDIA L20"
     python validate_apex_combo.py recommend --backend lammps --prefer cpu
 """
 
@@ -77,6 +77,7 @@
         "c8_m32_cpu",
     ],
     "lammps_gpu": [
+        "c16_m120_1 * NVIDIA L20",
         "c8_m32_1 * NVIDIA 4090",
         "c8_m31_1 * NVIDIA T4",
         "c4_m15_1 * NVIDIA T4",
@@ -200,7 +201,7 @@ def list_combos(backend: str = "lammps", prefer: str = "cpu") -> dict:
         "notes": [
             "Always validate image×scass before writing global.json / submitting.",
             "Outer Bohrium job should use c1_m2_cpu, never GPU.",
-            "DPA4 image 0.0.2 is validated on RTX 4090 for LAMMPS, phonon, and Grüneisen.",
+            "DPA4 image 0.0.2 is validated on NVIDIA L20 and RTX 4090; L20 is the default GPU resource.",
             "DPA4 image 0.0.2 is blocked on CPU machines after sequential deployment timeouts.",
             "For triclinic cells, avoid deepmd-kit:3.1.1; prefer the DPA4 0.0.2 image.",
         ],