refactor(recipe)!: backport vrnvl72 rename to r0.7.0 (#6419)

scripts/performance/README.md

@@ -239,11 +239,11 @@ Mounting cached files is not enough by itself. If `HF_HUB_OFFLINE` remains `0`,
 
 ##### Performance arguments
 
-- `-g/--gpu`: Target GPU type (`h100`, `b200`, `gb200`, `gb300`, `b300`).
+- `-g/--gpu`: Target GPU type (`h100`, `b200`, `gb200`, `gb300`, `b300`, `vrnvl72`). `vr200` is a deprecated alias for `vrnvl72`.
 - `-c/--compute_dtype`: Compute precision (`bf16`, `fp8_cs`, `fp8_mx`, `fp8_sc`, `nvfp4`). Default `bf16`.
 - `-vb/--enable_vboost`: Enable VBoost (tensor core power steering). Pass `true` or `false`. Disabled by default.
 - `-lgc/--lock_gpu_freq`: Lock GPU graphics clock to a fixed frequency in MHz (e.g. `1200`). Used for silicon simulation correlation studies. Disabled by default.
-- `-lmc/--peak_mem_clk`: Lock GPU memory clock to a fixed peak frequency in MHz (e.g. `2600`). Used for silicon simulation correlation studies. Defaults to `4752` MHz for VR200 and is disabled by default for other GPUs. Pass `-lmc -1` or `--peak_mem_clk -1` to disable the VR200 default.
+- `-lmc/--peak_mem_clk`: Lock GPU memory clock to a fixed peak frequency in MHz (e.g. `2600`). Used for silicon simulation correlation studies. Defaults to `4752` MHz for VRNVL72 and is disabled by default for other GPUs. Pass `-lmc -1` or `--peak_mem_clk -1` to disable the VRNVL72 default.
 - `-en/--enable_nsys`: Enable Nsight Systems profiling. Disabled by default.
 - `-pyp/--pytorch_profiler`: Enable PyTorch profiler. Pass `true` or `false`. Disabled by default.
 - `--profiling_start_step`: Defines start step for profiling. Default `10`.

scripts/performance/argument_parser.py

@@ -33,10 +33,14 @@
     "b300": 8,
     "gb200": 4,
     "gb300": 4,
-    "vr200": 4,
+    "vrnvl72": 4,
     "r100": 1,
 }
 
+# Deprecated --gpu names and their replacements. When the deprecation window ends, drop an entry together with
+# its counterpart in LEGACY_HARDWARE_ALIASES (scripts/training/recipe_metadata.py).
+LEGACY_GPU_ALIASES = {"vr200": "vrnvl72"}
+
 
 def list_of_strings(arg):
     """Split a comma-separated string into a list of substrings."""
@@ -80,6 +84,16 @@ def lower_str(arg):
     return arg.lower()
 
 
+def canonical_gpu(arg: str) -> str:
+    """Map a deprecated GPU name to its replacement before argparse checks the choices."""
+    if arg in LEGACY_GPU_ALIASES:
+        # Every rank re-parses the forwarded argv, so only the launcher and global rank 0 warn.
+        if os.environ.get("RANK", os.environ.get("SLURM_PROCID", "0")) == "0":
+            logger.warning("--gpu %s is deprecated; use --gpu %s.", arg, LEGACY_GPU_ALIASES[arg])
+        return LEGACY_GPU_ALIASES[arg]
+    return arg
+
+
 def bool_arg(arg):
     """Convert a string CLI value to a boolean."""
     if arg.lower() in ["true", "1", "t", "yes", "y"]:
@@ -787,9 +801,9 @@ def parse_cli_args():
     performance_args.add_argument(
         "-g",
         "--gpu",
-        type=str,
+        type=canonical_gpu,
         choices=NUM_GPUS_PER_NODE_MAP.keys(),
-        help="Target gpu type.",
+        help="Target gpu type. vr200 is a deprecated alias for vrnvl72.",
         required=True,
     )
     performance_args.add_argument(
@@ -831,8 +845,8 @@ def parse_cli_args():
         "--peak_mem_clk",
         help="Lock GPU memory clock to the specified peak frequency in MHz via "
         "`sudo nvidia-smi -lmc <freq>,<freq>`. Runs once per node before training. "
-        "Defaults to 4752 MHz for VR200 and is disabled by default for other GPUs. "
-        "Pass `-lmc -1` or `--peak_mem_clk -1` to disable the VR200 default.",
+        "Defaults to 4752 MHz for VRNVL72 and is disabled by default for other GPUs. "
+        "Pass `-lmc -1` or `--peak_mem_clk -1` to disable the VRNVL72 default.",
         type=int,
         required=False,
         default=None,

scripts/performance/dump_perf_configs.py

@@ -65,10 +65,10 @@
     ("llama", "llama3_8b", "pretrain", 8, "gb200", "nvfp4"),
     ("llama", "llama3_8b", "pretrain", 32, "gb200", "bf16"),
     ("llama", "llama3_8b", "pretrain", 32, "gb200", "fp8_cs"),
-    ("llama", "llama3_8b", "pretrain", 8, "vr200", "bf16"),
-    ("llama", "llama3_8b", "pretrain", 8, "vr200", "fp8_cs"),
-    ("llama", "llama3_8b", "pretrain", 8, "vr200", "fp8_mx"),
-    ("llama", "llama3_8b", "pretrain", 8, "vr200", "nvfp4"),
+    ("llama", "llama3_8b", "pretrain", 8, "vrnvl72", "bf16"),
+    ("llama", "llama3_8b", "pretrain", 8, "vrnvl72", "fp8_cs"),
+    ("llama", "llama3_8b", "pretrain", 8, "vrnvl72", "fp8_mx"),
+    ("llama", "llama3_8b", "pretrain", 8, "vrnvl72", "nvfp4"),
     ("llama", "llama3_8b", "pretrain", 8, "b300", "bf16"),
     ("llama", "llama3_8b", "pretrain", 8, "b300", "fp8_cs"),
     ("llama", "llama3_8b", "pretrain", 8, "b300", "fp8_mx"),
@@ -172,10 +172,10 @@
     ("deepseek", "deepseek_v3", "pretrain", 256, "gb200", "fp8_cs"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "gb200", "fp8_mx"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "gb200", "nvfp4"),
-    ("deepseek", "deepseek_v3", "pretrain", 256, "vr200", "bf16"),
-    ("deepseek", "deepseek_v3", "pretrain", 256, "vr200", "fp8_cs"),
-    ("deepseek", "deepseek_v3", "pretrain", 256, "vr200", "fp8_mx"),
-    ("deepseek", "deepseek_v3", "pretrain", 256, "vr200", "nvfp4"),
+    ("deepseek", "deepseek_v3", "pretrain", 256, "vrnvl72", "bf16"),
+    ("deepseek", "deepseek_v3", "pretrain", 256, "vrnvl72", "fp8_cs"),
+    ("deepseek", "deepseek_v3", "pretrain", 256, "vrnvl72", "fp8_mx"),
+    ("deepseek", "deepseek_v3", "pretrain", 256, "vrnvl72", "nvfp4"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "b300", "bf16"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "b300", "fp8_cs"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "b300", "fp8_mx"),
@@ -255,8 +255,8 @@
     ("kimi", "kimi_k2", "pretrain", 256, "gb200", "bf16"),
     ("kimi", "kimi_k2", "pretrain", 256, "gb200", "fp8_cs"),
     ("kimi", "kimi_k2", "pretrain", 256, "gb200", "fp8_mx"),
-    ("kimi", "kimi_k2", "pretrain", 256, "vr200", "bf16"),
-    ("kimi", "kimi_k2", "pretrain", 256, "vr200", "fp8_mx"),
+    ("kimi", "kimi_k2", "pretrain", 256, "vrnvl72", "bf16"),
+    ("kimi", "kimi_k2", "pretrain", 256, "vrnvl72", "fp8_mx"),
     ("kimi", "kimi_k2", "pretrain", 256, "b300", "bf16"),
     ("kimi", "kimi_k2", "pretrain", 256, "b300", "fp8_cs"),
     ("kimi", "kimi_k2", "pretrain", 256, "b300", "fp8_mx"),
@@ -321,9 +321,9 @@
     ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "gb300", "nvfp4"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 72, "gb300", "nvfp4"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 512, "gb300", "fp8_mx"),
-    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vr200", "nvfp4"),
-    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vr200", "fp8_mx"),
-    ("gpt_oss", "gpt_oss_20b", "pretrain", 64, "vr200", "nvfp4"),
+    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vrnvl72", "nvfp4"),
+    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vrnvl72", "fp8_mx"),
+    ("gpt_oss", "gpt_oss_20b", "pretrain", 64, "vrnvl72", "nvfp4"),
     # GPT-OSS 120B
     ("gpt_oss", "gpt_oss_120b", "pretrain", 64, "gb300", "bf16"),
     ("gpt_oss", "gpt_oss_120b", "pretrain", 64, "gb300", "fp8_mx"),
@@ -408,10 +408,10 @@
     ("qwen_vl", "qwen35_vl_397b_a17b", "pretrain", 256, "h100", "bf16"),
     ("qwen_vl", "qwen35_vl_397b_a17b", "pretrain", 256, "h100", "fp8_cs"),
     # Additional flattened coverage
-    ("deepseek", "deepseek_v3", "pretrain", 128, "vr200", "bf16"),
-    ("deepseek", "deepseek_v3", "pretrain", 128, "vr200", "fp8_cs"),
-    ("deepseek", "deepseek_v3", "pretrain", 128, "vr200", "fp8_mx"),
-    ("deepseek", "deepseek_v3", "pretrain", 128, "vr200", "nvfp4"),
+    ("deepseek", "deepseek_v3", "pretrain", 128, "vrnvl72", "bf16"),
+    ("deepseek", "deepseek_v3", "pretrain", 128, "vrnvl72", "fp8_cs"),
+    ("deepseek", "deepseek_v3", "pretrain", 128, "vrnvl72", "fp8_mx"),
+    ("deepseek", "deepseek_v3", "pretrain", 128, "vrnvl72", "nvfp4"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "gb300", "fp8_mx", "large_scale"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "gb200", "fp8_mx", "large_scale"),
     ("deepseek", "deepseek_v3", "pretrain", 256, "b300", "fp8_mx", "large_scale"),
@@ -421,18 +421,18 @@
     ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "gb200", "nvfp4"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "b300", "fp8_mx"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "b300", "nvfp4"),
-    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vr200", "fp8_mx"),
-    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vr200", "nvfp4"),
+    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vrnvl72", "fp8_mx"),
+    ("gpt_oss", "gpt_oss_20b", "pretrain", 8, "vrnvl72", "nvfp4"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 64, "b300", "fp8_mx"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 64, "b300", "nvfp4"),
-    ("gpt_oss", "gpt_oss_20b", "pretrain", 64, "vr200", "nvfp4"),
+    ("gpt_oss", "gpt_oss_20b", "pretrain", 64, "vrnvl72", "nvfp4"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 72, "gb300", "nvfp4"),
     ("gpt_oss", "gpt_oss_20b", "pretrain", 72, "gb200", "nvfp4"),
-    ("gpt_oss", "gpt_oss_120b", "pretrain", 64, "vr200", "bf16"),
-    ("gpt_oss", "gpt_oss_120b", "pretrain", 64, "vr200", "fp8_mx"),
-    ("llama", "llama3_70b", "pretrain", 64, "vr200", "bf16"),
-    ("llama", "llama3_70b", "pretrain", 64, "vr200", "fp8_mx"),
-    ("llama", "llama3_70b", "pretrain", 64, "vr200", "nvfp4"),
+    ("gpt_oss", "gpt_oss_120b", "pretrain", 64, "vrnvl72", "bf16"),
+    ("gpt_oss", "gpt_oss_120b", "pretrain", 64, "vrnvl72", "fp8_mx"),
+    ("llama", "llama3_70b", "pretrain", 64, "vrnvl72", "bf16"),
+    ("llama", "llama3_70b", "pretrain", 64, "vrnvl72", "fp8_mx"),
+    ("llama", "llama3_70b", "pretrain", 64, "vrnvl72", "nvfp4"),
     ("llama", "llama31_405b", "pretrain", 128, "gb300", "bf16"),
     ("llama", "llama31_405b", "pretrain", 128, "gb300", "fp8_cs"),
     ("llama", "llama31_405b", "pretrain", 128, "gb300", "fp8_mx"),
@@ -447,19 +447,19 @@
     ("llama", "llama31_405b", "pretrain", 256, "b300", "nvfp4"),
     ("llama", "llama31_405b", "pretrain", 256, "b200", "fp8_mx"),
     ("llama", "llama31_405b", "pretrain", 256, "b200", "nvfp4"),
-    ("llama", "llama31_405b", "pretrain", 256, "vr200", "bf16"),
-    ("llama", "llama31_405b", "pretrain", 256, "vr200", "fp8_mx"),
-    ("llama", "llama31_405b", "pretrain", 256, "vr200", "nvfp4"),
-    ("nemotronh", "nemotron_3_nano", "pretrain", 8, "vr200", "bf16"),
-    ("nemotronh", "nemotron_3_nano", "pretrain", 8, "vr200", "fp8_mx"),
-    ("nemotronh", "nemotron_3_nano", "pretrain", 8, "vr200", "nvfp4"),
-    ("nemotronh", "nemotron_3_super", "pretrain", 64, "vr200", "bf16"),
-    ("nemotronh", "nemotron_3_super", "pretrain", 64, "vr200", "fp8_mx"),
-    ("nemotronh", "nemotron_3_super", "pretrain", 64, "vr200", "nvfp4"),
-    ("nemotronh", "nemotron_3_ultra", "pretrain", 256, "vr200", "fp8_mx"),
-    ("nemotronh", "nemotron_3_ultra", "pretrain", 256, "vr200", "nvfp4"),
-    ("qwen", "qwen3_30b_a3b", "pretrain", 8, "vr200", "bf16"),
-    ("qwen", "qwen3_30b_a3b", "pretrain", 8, "vr200", "fp8_mx"),
+    ("llama", "llama31_405b", "pretrain", 256, "vrnvl72", "bf16"),
+    ("llama", "llama31_405b", "pretrain", 256, "vrnvl72", "fp8_mx"),
+    ("llama", "llama31_405b", "pretrain", 256, "vrnvl72", "nvfp4"),
+    ("nemotronh", "nemotron_3_nano", "pretrain", 8, "vrnvl72", "bf16"),
+    ("nemotronh", "nemotron_3_nano", "pretrain", 8, "vrnvl72", "fp8_mx"),
+    ("nemotronh", "nemotron_3_nano", "pretrain", 8, "vrnvl72", "nvfp4"),
+    ("nemotronh", "nemotron_3_super", "pretrain", 64, "vrnvl72", "bf16"),
+    ("nemotronh", "nemotron_3_super", "pretrain", 64, "vrnvl72", "fp8_mx"),
+    ("nemotronh", "nemotron_3_super", "pretrain", 64, "vrnvl72", "nvfp4"),
+    ("nemotronh", "nemotron_3_ultra", "pretrain", 256, "vrnvl72", "fp8_mx"),
+    ("nemotronh", "nemotron_3_ultra", "pretrain", 256, "vrnvl72", "nvfp4"),
+    ("qwen", "qwen3_30b_a3b", "pretrain", 8, "vrnvl72", "bf16"),
+    ("qwen", "qwen3_30b_a3b", "pretrain", 8, "vrnvl72", "fp8_mx"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 64, "gb300", "bf16"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 64, "gb300", "fp8_cs"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 64, "gb300", "fp8_mx"),
@@ -476,9 +476,9 @@
     ("qwen", "qwen3_235b_a22b", "pretrain", 64, "b200", "fp8_cs"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 64, "b200", "fp8_mx"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 64, "b200", "nvfp4"),
-    ("qwen", "qwen3_235b_a22b", "pretrain", 256, "vr200", "bf16"),
-    ("qwen", "qwen3_235b_a22b", "pretrain", 256, "vr200", "fp8_mx"),
-    ("qwen", "qwen3_235b_a22b", "pretrain", 256, "vr200", "nvfp4"),
+    ("qwen", "qwen3_235b_a22b", "pretrain", 256, "vrnvl72", "bf16"),
+    ("qwen", "qwen3_235b_a22b", "pretrain", 256, "vrnvl72", "fp8_mx"),
+    ("qwen", "qwen3_235b_a22b", "pretrain", 256, "vrnvl72", "nvfp4"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 256, "gb300", "fp8_mx", "large_scale"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 256, "gb200", "fp8_mx", "large_scale"),
     ("qwen", "qwen3_235b_a22b", "pretrain", 256, "b300", "fp8_mx", "large_scale"),

scripts/performance/setup_experiment.py

@@ -672,7 +672,7 @@ def main(
         ]
     )
 
-    if peak_mem_clk is None and gpu == "vr200":
+    if peak_mem_clk is None and gpu == "vrnvl72":
         peak_mem_clk = 4752
     if peak_mem_clk == -1:
         peak_mem_clk = None

scripts/performance/utils/executors.py

@@ -185,7 +185,7 @@ def slurm_executor(
     log_repo_status_cmd = "bash /opt/Megatron-Bridge/docker/common/print_sha.sh /nemo_run/configs/repo_status.json"
     custom_bash_cmds.append(log_repo_status_cmd)
 
-    numa_divisor = 2 if gpu.lower() in ["gb200", "gb300", "vr200"] else 4
+    numa_divisor = 2 if gpu.lower() in ["gb200", "gb300", "vrnvl72"] else 4
     numa_cmd = f"numactl --cpunodebind=$((SLURM_LOCALID/{numa_divisor})) --membind=$((SLURM_LOCALID/{numa_divisor}))"
     if gpu.lower() in ["b300"] and enable_pct_binding:
         numa_cmd += " -C $((SLURM_LOCALID * 16)),$((SLURM_LOCALID * 16 + 1))"

scripts/performance/utils/overrides.py

@@ -280,7 +280,7 @@ def _apply_flat_cli_environment_compatibility(
     if connection_override:
         moe_a2a_overlap = args.moe_a2a_overlap if args.moe_a2a_overlap is not None else base_moe_a2a_overlap
         max_connections = 32 if effective_dispatcher_backend in {"deepep", "hybridep", "ncclep"} else 8
-        if gpu in {"b200", "b300", "gb200", "gb300", "vr200"}:
+        if gpu in {"b200", "b300", "gb200", "gb300", "vrnvl72"}:
             max_connections = 32
         elif (tp_size > 1 or cp_size > 1) and not moe_a2a_overlap:
             max_connections = 1
@@ -292,7 +292,7 @@ def _apply_flat_cli_environment_compatibility(
         if gpu in {"h100", "b200", "b300"}:
             domain_size = 8
             ranks_per_domain = min(ep_size, domain_size)
-        elif gpu in {"gb200", "gb300", "vr200", "r100"}:
+        elif gpu in {"gb200", "gb300", "vrnvl72", "r100"}:
             domain_size = 72
             if ep_size > domain_size:
                 raise ValueError("HybridEP expert parallel size must not exceed the 72-rank NVLink domain.")

scripts/performance/utils/utils.py

@@ -324,7 +324,7 @@ def apply_target_topology_environment(
             "NVLINK_DOMAIN_SIZE": 8,
             "USE_MNNVL": 0,
         }
-    elif normalized_gpu in {"gb200", "gb300", "vr200", "r100"}:
+    elif normalized_gpu in {"gb200", "gb300", "vrnvl72", "r100"}:
         if expert_model_parallel_size > 72:
             raise ValueError("HybridEP expert parallel size must not exceed the 72-rank NVLink domain.")
         topology = {

scripts/training/README.md

@@ -107,6 +107,9 @@ through configurable library recipes. Explicit benchmark dataset replacement rem
 Five legacy duplicate names resolve to the benchmark definition; their library workloads remain available through
 the corresponding generic recipe aliases. New recipe names should be unique across both packages.
 
+Benchmark recipe names that use the deprecated `vr200` hardware name resolve to the renamed `vrnvl72` recipe with a
+warning.
+
 Text recipes default to `llm_step`; all recipes infer their modality-specific forward step from the same registry.
 Pass `--step-func NAME` to override that selection. Common training,
 sequence-length, parallelism, optimization, and checkpoint fields also have convenience flags such as
@@ -340,8 +343,8 @@ populate the executor's `additional_parameters` and are not forwarded to the
 training command or its srun step.
 
 Use `-lmc FREQ`, `--peak-mem-clk FREQ`, or `--peak_mem_clk FREQ` to lock the GPU memory clock to a fixed frequency in
-MHz once per node before training. VR200 benchmark recipes default to `4752`; other recipes leave the memory clock
-unlocked. Pass `-lmc -1` to disable the VR200 default explicitly.
+MHz once per node before training. VRNVL72 benchmark recipes default to `4752`; other recipes leave the memory clock
+unlocked. Pass `-lmc -1` to disable the VRNVL72 default explicitly.
 
 The compact launcher does not add rank-command prefixes or infer target-specific
 allocation and NUMA policies. Configure those explicitly for the target.

scripts/training/recipe_metadata.py

@@ -19,11 +19,15 @@
 import argparse
 import ast
 import functools
+import logging
+import os
 import re
 from dataclasses import dataclass
 from pathlib import Path
 
 
+logger = logging.getLogger(__name__)
+
 BENCHMARK_RECIPE_FAMILY_PREFIXES = (
     ("qwen3_vl_", "qwen_vl"),
     ("qwen35_vl_", "qwen_vl"),
@@ -45,6 +49,10 @@
     r"(?:_[a-z0-9_]+)?_config$"
 )
 
+# Deprecated recipe hardware names and their replacements. When the deprecation window ends, drop an entry together
+# with its counterpart in LEGACY_GPU_ALIASES (scripts/performance/argument_parser.py).
+LEGACY_HARDWARE_ALIASES = {"vr200": "vrnvl72"}
+
 BENCHMARK_RECIPE_ROOT = Path(__file__).resolve().parents[2] / "src" / "megatron" / "bridge" / "perf_recipes"
 
 # These names are exported by both recipe packages today. Bare recipe lookup
@@ -181,8 +189,31 @@ def available_benchmark_recipe_metadata(recipe_name: str) -> BenchmarkRecipeMeta
     return metadata
 
 
+@functools.lru_cache(maxsize=None)
+def _canonical_recipe_name(recipe_name: str) -> str:
+    """Replace a deprecated hardware name in a benchmark recipe name when the renamed recipe is exported.
+
+    Cached so each process logs the deprecation once per requested name.
+    """
+    match = BENCHMARK_RECIPE_PATTERN.search(recipe_name)
+    if match is None or match.group("hardware") not in LEGACY_HARDWARE_ALIASES:
+        return recipe_name
+    start, end = match.span("hardware")
+    canonical_name = f"{recipe_name[:start]}{LEGACY_HARDWARE_ALIASES[match.group('hardware')]}{recipe_name[end:]}"
+    if available_benchmark_recipe_metadata(canonical_name) is None:
+        return recipe_name
+    # Every rank re-runs this lookup, so only the launcher and global rank 0 warn.
+    if os.environ.get("RANK", os.environ.get("SLURM_PROCID", "0")) == "0":
+        logger.warning("Recipe '%s' uses a deprecated hardware name; use '%s'.", recipe_name, canonical_name)
+    return canonical_name
+
+
 def resolved_benchmark_recipe_metadata(recipe_name: str) -> BenchmarkRecipeMetadata | None:
-    """Return benchmark metadata after applying unified lookup precedence."""
+    """Return benchmark metadata after applying legacy hardware aliases and unified lookup precedence.
+
+    A deprecated hardware name resolves to the renamed recipe, so callers load ``metadata.recipe_name``.
+    """
+    recipe_name = _canonical_recipe_name(recipe_name)
     if recipe_name in LIBRARY_RECIPE_PRECEDENCE_COLLISIONS:
         return None
     return available_benchmark_recipe_metadata(recipe_name)

scripts/training/recipe_runner.py

@@ -233,7 +233,10 @@ def load_recipe(
     seq_length: int | None = None,
 ) -> ConfigContainer:
     """Load an exact benchmark recipe when exported, otherwise a library recipe."""
-    if resolved_benchmark_recipe_metadata(recipe_name) is not None:
+    benchmark_metadata = resolved_benchmark_recipe_metadata(recipe_name)
+    if benchmark_metadata is not None:
+        # The resolved name replaces a deprecated hardware name in the requested one.
+        recipe_name = benchmark_metadata.recipe_name
         config_builder = find_benchmark_recipe(recipe_name)
         package_name = "megatron.bridge.perf_recipes"
         recipe_kind = "benchmark"