zhouhelena1 · Oct 9, 2026 · Oct 9, 2026
diff --git a/modal_dojo/common/models/__init__.py b/modal_dojo/common/models/__init__.py
@@ -9,7 +9,7 @@
     parse_inkling_response,
     parse_qwen3_6_response,
 )
-from .deepseek_v41_flash import DeepSeek_V4_1_Flash
+from .deepseek_v4_1_flash import DeepSeek_V4_1_Flash
 from .gemma4_26b_a4b import Gemma4_26B_A4B
 from .glm_4_7 import GLM_4_7
 from .glm_5_3_flash import GLM_5_3_Flash, GLM_5_3_Flash_LoRA

diff --git a/..._dojo/common/models/deepseek_v41_flash.py → ...dojo/common/models/deepseek_v4_1_flash.py b/..._dojo/common/models/deepseek_v41_flash.py → ...dojo/common/models/deepseek_v4_1_flash.py
@@ -8,7 +8,7 @@
 Neither the indexer nor Engram is representable as a ``ModelArchitecture``, so
 ``architecture`` is left ``None`` and the Miles recipe renders upstream's
 ``scripts/models/deepseek-v4.1.py`` via ``miles_model_name`` — including its
-``--spec miles_plugins.models.deepseek_v41.deepseek_v41 get_dsv41_spec``. Same
+``--spec miles_plugins.models.deepseek_v4_1.deepseek_v4_1 get_dsv41_spec``. Same
 arrangement as ``Qwen3_5_4B_Miles``.
 
 The released checkpoint is fp8; Miles converts it to a bf16 torch_dist

diff --git a/modal_dojo/common/models/validation.py b/modal_dojo/common/models/validation.py
@@ -10,7 +10,7 @@
 
 from ..framework import Framework
 from .base import ModelConfig
-from .deepseek_v41_flash import DeepSeek_V4_1_Flash
+from .deepseek_v4_1_flash import DeepSeek_V4_1_Flash
 from .gemma4_26b_a4b import Gemma4_26B_A4B
 from .glm_4_7 import GLM_4_7
 from .glm_5_3_flash import GLM_5_3_Flash_LoRA

diff --git a/modal_dojo/frameworks/miles/modal_helpers/patches/patch_advantage_distribution.py b/modal_dojo/frameworks/miles/modal_helpers/patches/patch_advantage_distribution.py
@@ -76,8 +76,15 @@ def _patch_file(path: Path) -> None:
     print(f"Patched {path.name} with advantage-distribution reporting")
 
 
+# radixark/miles#3248 moved log_utils.py under training_utils/metrics/.
+_CANDIDATES = (
+    Path("/root/miles/miles/backends/training_utils/metrics/log_utils.py"),
+    Path("/root/miles/miles/backends/training_utils/log_utils.py"),
+)
+
+
 def main() -> None:
-    _patch_file(Path("/root/miles/miles/backends/training_utils/log_utils.py"))
+    _patch_file(next((p for p in _CANDIDATES if p.exists()), _CANDIDATES[-1]))
 
 
 if __name__ == "__main__":

diff --git a/...ches/patch_deepseek_v41_checksum_audit.py → ...hes/patch_deepseek_v4_1_checksum_audit.py b/...ches/patch_deepseek_v41_checksum_audit.py → ...hes/patch_deepseek_v4_1_checksum_audit.py
diff --git a/...ches/patch_deepseek_v41_checksum_cache.py → ...hes/patch_deepseek_v4_1_checksum_cache.py b/...ches/patch_deepseek_v41_checksum_cache.py → ...hes/patch_deepseek_v4_1_checksum_cache.py
@@ -5,8 +5,8 @@
 through cuMemCreate, which cannot reclaim that cache. Miles calls onload_kv
 immediately after the checksum, so return the scratch memory to CUDA first.
 
-Image: radixark/miles:dsv41-h200-ea751aac8
-SGLang: 7e74b31b2668934cf89dbe15188a010fda381aba
+Image: radixark/miles:dev-202610082133
+SGLang: b84bec5761a6cdc8b1df86b9b89ead80a7a9bb02
 """
 
 from pathlib import Path

diff --git a/...es/patch_deepseek_v41_checksum_offsets.py → ...s/patch_deepseek_v4_1_checksum_offsets.py b/...es/patch_deepseek_v41_checksum_offsets.py → ...s/patch_deepseek_v4_1_checksum_offsets.py
@@ -1,12 +1,12 @@
 """Use 64-bit offsets when SGLang hashes DeepSeek's large weight tensors.
 
-The H200 image's post-sync checksum audit hashes every model tensor. Its
+The image's post-sync checksum audit hashes every model tensor. Its
 Triton hash kernel multiplies a signed 32-bit program ID by the tile size,
 overflowing beyond 2**31 uint32 words (8 GiB) and reading invalid addresses.
 Promote before multiplication; the intentional uint32 hash mixing stays intact.
 
-Image: radixark/miles:dsv41-h200-ea751aac8
-SGLang: 7e74b31b2668934cf89dbe15188a010fda381aba
+Image: radixark/miles:dev-202610082133
+SGLang: b84bec5761a6cdc8b1df86b9b89ead80a7a9bb02
 """
 
 from pathlib import Path

diff --git a/...s/patch_deepseek_v41_fp4_dequant_block.py → .../patch_deepseek_v4_1_fp4_dequant_block.py b/...s/patch_deepseek_v41_fp4_dequant_block.py → .../patch_deepseek_v4_1_fp4_dequant_block.py
@@ -14,8 +14,8 @@
 
 Executed at image-build time via ``python3 <this file>``.
 
-Image: radixark/miles:dsv41-h200-ea751aac8
-SGLang: 7e74b31b2668934cf89dbe15188a010fda381aba
+Image: radixark/miles:dev-202610082133
+SGLang: b84bec5761a6cdc8b1df86b9b89ead80a7a9bb02
 """
 
 import pathlib

diff --git a/...patch_deepseek_v41_router_registration.py → ...atch_deepseek_v4_1_router_registration.py b/...patch_deepseek_v41_router_registration.py → ...atch_deepseek_v4_1_router_registration.py
@@ -1,4 +1,4 @@
-"""Wait for asynchronous router registration in the DeepSeek V4.1 image.
+"""Wait for asynchronous router registration for the DeepSeek V4.1 recipe.
 
 Miles' SGLangRouterApiClient treats HTTP 202 as completed registration. The
 router is still discovering engine metadata at that point; a following weight

diff --git a/...patch_deepseek_v41_vision_topk_capture.py → ...atch_deepseek_v4_1_vision_topk_capture.py b/...patch_deepseek_v41_vision_topk_capture.py → ...atch_deepseek_v4_1_vision_topk_capture.py
@@ -18,8 +18,8 @@
 
 Executed at image-build time via ``python3 <this file>``.
 
-Image: radixark/miles:dsv41-h200-ea751aac8
-SGLang: 7e74b31b2668934cf89dbe15188a010fda381aba
+Image: radixark/miles:dev-202610082133
+SGLang: b84bec5761a6cdc8b1df86b9b89ead80a7a9bb02
 """
 
 import pathlib

diff --git a/modal_dojo/frameworks/miles/modal_helpers/patches/patch_zero_std_metrics.py b/modal_dojo/frameworks/miles/modal_helpers/patches/patch_zero_std_metrics.py
@@ -2,16 +2,27 @@
 
 from pathlib import Path
 
-OLD = "str(round(g[0].get_reward_value(args), 1))"
-NEW = "str(round(float(g[0].get_reward_value(args)), 1))"
+# radixark/miles#3722 hoisted the reward into ``uniform_rewards`` but still
+# buckets an int 1 as "1" and a float 1.0 as "1.0".
+_ANCHORS = (
+    (
+        "str(round(g[0].get_reward_value(args), 1))",
+        "str(round(float(g[0].get_reward_value(args)), 1))",
+    ),
+    (
+        "str(round(reward, 1)) for reward in uniform_rewards",
+        "str(round(float(reward), 1)) for reward in uniform_rewards",
+    ),
+)
 
 
 def patch_source(source: str) -> str:
-    if NEW in source:
-        return source
-    if source.count(OLD) != 1:
-        raise ValueError("Miles zero-std reward bucket anchor changed")
-    return source.replace(OLD, NEW, 1)
+    for old, new in _ANCHORS:
+        if new in source:
+            return source
+        if source.count(old) == 1:
+            return source.replace(old, new, 1)
+    raise ValueError("Miles zero-std reward bucket anchor changed")
 
 
 def main() -> None:

diff --git a/modal_dojo/train_recipes/miles_recipe/__init__.py b/modal_dojo/train_recipes/miles_recipe/__init__.py
@@ -1,5 +1,5 @@
 from modal_dojo.train_recipes.miles_recipe.recipe import MilesRecipe
-from modal_dojo.train_recipes.miles_recipe.deepseek_v41_flash import (
+from modal_dojo.train_recipes.miles_recipe.deepseek_v4_1_flash import (
     DeepSeek_V4_1_Flash_Recipe,
 )
 from modal_dojo.train_recipes.miles_recipe.gemma4_26b_a4b import (

diff --git a/...ecipes/miles_recipe/deepseek_v41_flash.py → ...cipes/miles_recipe/deepseek_v4_1_flash.py b/...ecipes/miles_recipe/deepseek_v41_flash.py → ...cipes/miles_recipe/deepseek_v4_1_flash.py
@@ -1,4 +1,4 @@
-"""DeepSeek-V4.1-Flash GRPO recipe, ported from upstream's ``run_deepseek_v41.py``."""
+"""DeepSeek-V4.1-Flash GRPO recipe, ported from upstream's ``run_deepseek_v4_1.py``."""
 
 from dataclasses import field
 from pathlib import Path
@@ -11,11 +11,10 @@
 from modal_dojo.common.patches import encode_patch
 from modal_dojo.train_recipes.miles_recipe.recipe import MilesRecipe
 
-# The dedicated amd64/H200 image includes the DeepSeek sources together:
-# miles ea751aac8, sglang 7e74b31b2, and Megatron-LM b0b23e198.
-# Keep those bundled revisions instead of overlaying the older PR head and
-# the arm64 deepseek-v41 image's SGLang tree. The named tag is outside miles'
-# automatic dev/PR tag cleanup; the digest also prevents tag updates drifting.
+# The standard multi-arch dev image carries DeepSeek-V4.1 on miles main:
+# miles f5757caab (with the radixark/miles#3605-#3634 kernel stack),
+# sglang-miles b84bec576, and Megatron-LM fd15ee20a. The digest pins it, since
+# miles prunes old dated dev tags and the tag could otherwise drift.
 
 # Node-local disk: 768 GiB stages the params-only save (~1.1 TB bf16 over 8
 # nodes, so ~140 GB a node); 1.5 TiB holds the NVMe-streamed optimizer
@@ -43,12 +42,12 @@
 
 # Build-time shims for gaps in the bundled sources; see each script's docstring.
 _PATCHES = (
-    "patch_deepseek_v41_router_registration",
-    "patch_deepseek_v41_checksum_audit",
-    "patch_deepseek_v41_checksum_cache",
-    "patch_deepseek_v41_checksum_offsets",
-    "patch_deepseek_v41_fp4_dequant_block",
-    "patch_deepseek_v41_vision_topk_capture",
+    "patch_deepseek_v4_1_router_registration",
+    "patch_deepseek_v4_1_checksum_audit",
+    "patch_deepseek_v4_1_checksum_cache",
+    "patch_deepseek_v4_1_checksum_offsets",
+    "patch_deepseek_v4_1_fp4_dequant_block",
+    "patch_deepseek_v4_1_vision_topk_capture",
 )
 
 
@@ -66,8 +65,8 @@ class DeepSeek_V4_1_Flash_Recipe(MilesRecipe):
     model_config_class: ClassVar[type[ModelConfig]] = DeepSeek_V4_1_Flash
 
     docker_image: str = (
-        "radixark/miles:dsv41-h200-ea751aac8@sha256:"
-        "eff12eb317af7021f637ab6bf21359164e27274ca07a3b2b3754dfa4c5d1bd66"
+        "radixark/miles:dev-202610082133@sha256:"
+        "9cf2ed11d587af9884c75344681cf4100b8025eb357cb4d09a063823d4d81975"
     )
     image_run_commands: list[str] = field(default_factory=_image_patches)
     gpu_type: str = "H200"
@@ -80,7 +79,7 @@ class DeepSeek_V4_1_Flash_Recipe(MilesRecipe):
     # ${MODEL_ARGS[@]} verbatim — including its --spec for the V4.1 layer spec.
     miles_model_name: str = "deepseek-v4.1"
     # Selects miles' megatron→HF weight mapping (miles/backends/megatron_utils/
-    # megatron_to_hf/deepseekv41.py).
+    # megatron_to_hf/deepseekv4_1.py).
     model_name: str = "deepseekv41"
     # Selects the miles (not sglang) V4.1 training implementation.
     dsv4_impl: str = "miles"
@@ -115,6 +114,9 @@ class DeepSeek_V4_1_Flash_Recipe(MilesRecipe):
             "SGLANG_HEALTH_CHECK_TIMEOUT": "900",
             "SGLANG_DG_CACHE_DIR_PER_PROCESS": "1",
             "SGLANG_OPT_FP8_WO_A_GEMM": "0",
+            # Upstream: the compensated mHC weight split caches derived weights
+            # that online weight updates cannot refresh.
+            "SGLANG_OPT_DEEPGEMM_HC_PRENORM": "0",
             "SGLANG_OPT_FUSE_WQA_WKV": "0",
             "SGLANG_DISABLE_MULTIMEM_AG": "1",
             # Frozen Engram tables are omitted from weight sync. Keep them in
@@ -132,7 +134,7 @@ class DeepSeek_V4_1_Flash_Recipe(MilesRecipe):
 
     # ── Checkpoints ──────────────────────────────────────────────────────────
     # "raw" turns on the HF → torch_dist conversion that upstream runs as
-    # `run_deepseek_v41.py prepare-spmd`; miles then loads it as ref_load.
+    # `run_deepseek_v4_1.py prepare-spmd`; miles then loads it as ref_load.
     megatron_to_hf_mode: str = "raw"
     ref_load: str = "/checkpoints/DeepSeek-V4.1-Flash_torch_dist"
 

diff --git a/modal_dojo/train_recipes/miles_recipe/recipe.py b/modal_dojo/train_recipes/miles_recipe/recipe.py
@@ -913,7 +913,7 @@ def train_async(self) -> bool:
 
     @classmethod
     def get_base_recipe(cls, model_config: ModelConfig) -> "MilesRecipe | None":
-        from modal_dojo.train_recipes.miles_recipe.deepseek_v41_flash import (
+        from modal_dojo.train_recipes.miles_recipe.deepseek_v4_1_flash import (
             DeepSeek_V4_1_Flash_Recipe,
         )
         from modal_dojo.train_recipes.miles_recipe.gemma4_26b_a4b import (