freesolo-flash-dev 1.4.22__tar.gz → 1.4.23__tar.gz

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: freesolo-flash-dev
-Version: 1.4.22
+Version: 1.4.23
 Summary: Flash — managed LoRA post-training (SFT/GRPO) for Freesolo environments, driven by the `flash` CLI
 Project-URL: Homepage, https://github.com/freesolo-co/flash
 Project-URL: Repository, https://github.com/freesolo-co/flash

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/flash/core/catalog.py

@@ -376,7 +376,7 @@ MODELS: dict[str, ModelInfo] = {
         quant="bf16",
         recommended_gpu="H100",
         serving=ServingCapacity(
-            gpu="L40S",
+            gpu="H100",
             serve_model_id=SERVING_MODEL_REPOS["Qwen/Qwen3.5-9B"],
             max_loras=16,
             max_cpu_loras=16,
@@ -543,7 +543,7 @@ MODELS["Qwen/Qwen3.5-9B-Base"] = replace(
     display_name="Qwen3.5 9B Base (pretrained)",
     managed_revision="68c46c4b3498877f3ef123c856ecfde50c39f404",
     serving=ServingCapacity(
-        gpu="L40S",
+        gpu="H100",
         serve_model_id="Qwen/Qwen3.5-9B-Base",
         max_loras=16,
         max_cpu_loras=16,
@@ -557,7 +557,7 @@ MODELS["Qwen/Qwen3.5-9B-Base"] = replace(
     notes="Pretrained-only Qwen3.5 9B; bf16 LoRA SFT/GRPO. Uses its own weights and "
     "tokenizer, not the post-trained 9B. Native chat/thinking formatting is available for "
     "fine-tuning; instruction following and reasoning are not post-trained. Hosted FP8 serving "
-    "on L40S; customer-owned BF16 serving on H100.",
+    "on H100; customer-owned BF16 serving on H100.",
 )
 
 

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/flash/providers/nebius/api.py

@@ -181,10 +181,15 @@ def cloud_init(payload, capsule_ref, image, *, deadline_at):
         "--workdir",
         "/root/flash",
         image,
-        "python",
-        "-u",
-        "/opt/flash/capsule.pyz",
-        "bootstrap",
+        "sh",
+        "-c",
+        # Qualified image digests can predate the Azure artifact backend. Match
+        # Modal's launch-time dependencies before the capsule publishes liveness.
+        (
+            "python -m pip install --disable-pip-version-check --no-input "
+            "'azure-storage-blob>=12.24,<13' 'azure-identity>=1.19,<2' "
+            "&& exec python -u /opt/flash/capsule.pyz bootstrap"
+        ),
     ]
     user_data = "#cloud-config\n" + json.dumps(
         {

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/flash/providers/nebius/bootstrap.py

@@ -17,9 +17,12 @@ def publish_capsule(payload, capsule, *, deadline_at):
     blob = f"{repo_prefix(payload['hf_repo'])}/capsules/{digest}.pyz"
 
     def publish():
-        with open_container(token) as container:
+        # Container properties require control-plane credentials; a worker
+        # user-delegation SAS only authorizes blob operations.
+        with open_container() as container:
             if container.get_container_properties().public_access:
                 raise ValueError("Nebius capsule container must be private")
+        with open_container(token) as container:
             BlobArtifacts(container).immutable(blob, capsule)
 
     artifact_call(

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/flash/server/routes/serving_chat.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 from typing import Any
 
+import anyio
 import httpx
 from fastapi import HTTPException
 from fastapi.responses import StreamingResponse
@@ -74,7 +75,10 @@ class _UpstreamStreamingResponse(StreamingResponse):
         try:
             await super().__call__(scope, receive, send)
         finally:
-            self._upstream.close()
+            # Closing a synchronous HTTP stream can wait on transport locks or I/O.
+            # Keep health checks responsive, and finish cleanup even on disconnect.
+            with anyio.CancelScope(shield=True):
+                await anyio.to_thread.run_sync(self._upstream.close)
 
 
 def _resolve_chat_request(

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/flash/serving/src/engine/model_config.py

@@ -38,7 +38,7 @@ from flash.serving.src.engine.prequant_config import (
 # checkpoints published to the operator HF org.
 #
 # sizing rationale. preallocated lora buffers and loaded checkpoint weights dominate engine vram.
-# The dense 0.8B uses L4, the BF16 4B and both FP8 9B models use L40S, the FP8 27B uses H100, and
+# The dense 0.8B uses L4, the BF16 4B uses L40S, both FP8 9B models and the FP8 27B use H100, and
 # the BF16 35B MoE uses H200.
 #   - Qwen3.6-35B-A3B (vision-language MoE; arch ``Qwen3_5MoeForConditionalGeneration``) -> H200
 #     (141 GB) with the base bf16 weights, 6 x 64 LoRA at native 256k. bf16 (not FP8) is the one path giving
@@ -93,7 +93,7 @@ SERVING_MODELS: list[dict[str, Any]] = [
     {
         "base_model": "Qwen/Qwen3.5-9B",
         "image_input_limit": MAX_IMAGES_PER_CONVERSATION,
-        "gpu": "L40S",
+        "gpu": "H100",
         "engine": {
             "tokenizer_model": "Qwen/Qwen3.5-9B",
             "tokenizer_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
@@ -106,8 +106,7 @@ SERVING_MODELS: list[dict[str, Any]] = [
             "max_loras": 16,
             "max_lora_rank": 128,  # rank-128 / 16 hot LoRAs (cheap on the 9 GiB FP8 9B); native 256k context.
             "max_model_len": 262144,
-            # L40S: 64 improves short/4k throughput ~127%/~40% over sixteen and
-            # completes a full 64-request 32k-prompt rank-128 q_proj LoRA burst.
+            # Keep the qualified 64-sequence scheduler capacity on H100.
             "max_num_seqs": 64,
             "enforce_eager": False,
             "reasoning_parser": "qwen3",
@@ -116,7 +115,7 @@ SERVING_MODELS: list[dict[str, Any]] = [
     {
         "base_model": "Qwen/Qwen3.5-9B-Base",
         "image_input_limit": MAX_IMAGES_PER_CONVERSATION,
-        "gpu": "L40S",
+        "gpu": "H100",
         "engine": {
             # FP8 conversion of the pinned Base weights; never alias the Instruct checkpoint.
             "model_revision": "b1a2cf5f5e7bb4ad786e206f834569be3a3d8aa9",

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "freesolo-flash-dev"
-version = "1.4.22"
+version = "1.4.23"
 description ="Flash — managed LoRA post-training (SFT/GRPO) for Freesolo environments, driven by the `flash` CLI"
 # The training and serving stack is validated on Python 3.11 and 3.12.
 requires-python = ">=3.11,<3.13"
@@ -210,7 +210,7 @@ artifacts = ["flash/py.typed"]
 # auto-published by .github/workflows/publish-dev.yml: a push to `dev` whose version isn't yet on
 # PyPI publishes it; pushes that don't change it find it already published and no-op. So cutting a
 # dev-channel release = bump this (and [project].version) and merge to `dev`.
-version = "1.4.22"
+version = "1.4.23"
 
 [tool.pytest.ini_options]
 pythonpath = ["."]

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/serving/test_model_config.py

@@ -123,13 +123,13 @@ def test_no_bitsandbytes_anywhere() -> None:
         assert entry.get("load_format") != "bitsandbytes"
 
 
-def test_9b_has_l40s_rank128_serving_overrides() -> None:
+def test_9b_has_h100_rank128_serving_overrides() -> None:
     ov = engine_overrides_for("Qwen/Qwen3.5-9B")
     assert ov["max_loras"] == 16
     assert ov["max_lora_rank"] == 128
     assert ov["max_model_len"] == 262144
     assert ov["max_num_seqs"] == 64
-    assert gpu_for("Qwen/Qwen3.5-9B") == "L40S"
+    assert gpu_for("Qwen/Qwen3.5-9B") == "H100"
     assert ov["gpu_memory_utilization"] == 0.90
     assert (
         ov["enforce_eager"] is False

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/serving/test_per_model_gpu.py

@@ -471,8 +471,8 @@ def test_engine_concurrency_rejects_malformed_catalog_values(modal_app_module, m
     [
         ("Qwen/Qwen3.5-0.8B", "L4", 64),
         ("Qwen/Qwen3.5-4B", "L40S", 8),
-        ("Qwen/Qwen3.5-9B", "L40S", 64),
-        ("Qwen/Qwen3.5-9B-Base", "L40S", 64),
+        ("Qwen/Qwen3.5-9B", "H100", 64),
+        ("Qwen/Qwen3.5-9B-Base", "H100", 64),
         ("Qwen/Qwen3.8-27B", "H100", 8),
         ("Qwen/Qwen3.6-35B-A3B", "H200", 8),
     ],
@@ -544,7 +544,7 @@ def test_health_reports_pinned_gpu_over_derived_tier(modal_app_module):
         base_model = "Qwen/Qwen3.5-9B"
         registry = type("R", (), {"list_ready": lambda self: []})()
 
-    assert impl._health(_Bare())["configured_gpu"] == "L40S"
+    assert impl._health(_Bare())["configured_gpu"] == "H100"
 
 
 def test_health_reports_effective_max_model_len_override(modal_app_module):

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/serving/test_router.py

@@ -216,8 +216,8 @@ def test_healthz_reports_one_gpu_per_base_model(app_setup):
     ]
     assert body["base_models"] == [QWEN, QWEN_35B]  # sorted by model id
     assert body["gpus"] == 2  # two configured supported base-model engines
-    assert body["gpu_by_model"] == {QWEN: "L40S", QWEN_35B: "H200"}
-    assert body["gpu_tiers"] == ["H200", "L40S"]  # sorted distinct serving tiers
+    assert body["gpu_by_model"] == {QWEN: "H100", QWEN_35B: "H200"}
+    assert body["gpu_tiers"] == ["H100", "H200"]  # sorted distinct serving tiers
     assert "configuredGpu" not in body  # the single-GPU field is gone (per-model now)
     assert body["adapters"] == 3
 

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/test_catalog_consistency.py

@@ -58,7 +58,7 @@ def test_serving_capacity_matches_validated_matrix():
     # freesolo tree; compare both repos in the pinned freesolo-co/tests checkout.
     expected = {
         "Qwen/Qwen3.5-9B": {
-            "gpu": "L40S",
+            "gpu": "H100",
             "serve_model_id": "Freesolo-Co/Qwen3.5-9B-FP8",
             "max_loras": 16,
             "max_cpu_loras": 16,
@@ -104,7 +104,7 @@ def test_serving_capacity_matches_validated_matrix():
 
 def test_public_rows_include_serving_capacity():
     row = get_model("Qwen/Qwen3.5-9B").to_dict()
-    assert row["serving"]["gpu"] == "L40S"
+    assert row["serving"]["gpu"] == "H100"
     assert row["serving"]["max_loras"] == 16
     assert row["serving"]["max_lora_rank"] == 128
     assert row["serving"]["serve_model_id"] == "Freesolo-Co/Qwen3.5-9B-FP8"
@@ -114,7 +114,7 @@ def test_public_rows_prune_unset_serving_capacity_fields():
     row = get_model("Qwen/Qwen3.5-9B").to_dict()
     # serve_model_id survives while zero-valued optional capacity fields are pruned.
     assert row["serving"] == {
-        "gpu": "L40S",
+        "gpu": "H100",
         "serve_model_id": "Freesolo-Co/Qwen3.5-9B-FP8",
         "max_loras": 16,
         "max_cpu_loras": 16,

freesolo_flash_dev-1.4.23/tests/test_chat_stream_cleanup_responsiveness.py

@@ -0,0 +1,47 @@
+"""A stalled upstream close must not stall the control-plane event loop."""
+
+import asyncio
+import threading
+from types import SimpleNamespace
+
+import pytest
+from fastapi.responses import StreamingResponse
+
+from flash.server.routes.meta import health
+from flash.server.routes.serving_chat import _UpstreamStreamingResponse
+
+
+@pytest.mark.parametrize("disconnected", [False, True])
+def test_health_responds_while_upstream_close_waits(monkeypatch, disconnected):
+    entered = threading.Event()
+    release = threading.Event()
+    close_threads = []
+
+    def close():
+        close_threads.append(threading.get_ident())
+        entered.set()
+        assert release.wait(3), "event loop did not release stalled cleanup"
+
+    async def send_response(self, scope, receive, send):
+        if disconnected:
+            raise OSError("client disconnected")
+
+    monkeypatch.setattr(StreamingResponse, "__call__", send_response)
+
+    async def exercise():
+        response = _UpstreamStreamingResponse(iter(()), upstream=SimpleNamespace(close=close))
+        task = asyncio.create_task(response({}, None, None))
+        try:
+            assert await asyncio.wait_for(asyncio.to_thread(entered.wait, 1), 2)
+            assert (await asyncio.wait_for(health(), 0.2))["ok"]
+            assert len(close_threads) == 1
+            assert close_threads[0] != threading.get_ident()
+        finally:
+            release.set()
+            if disconnected:
+                with pytest.raises(OSError, match="client disconnected"):
+                    await task
+            else:
+                await task
+
+    asyncio.run(exercise())

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/test_nebius_spot.py

@@ -205,6 +205,13 @@ def test_create_always_attaches_server_price_cap_and_owned_disk(monkeypatch, dis
     assert spec["boot_disk"]["managed_disk"]["spec"]["type"] == "NETWORK_SSD_NON_REPLICATED"
     init = json.loads(spec["cloud_init_user_data"].split("\n", 1)[1])
     assert all(f["permissions"] == "0600" for f in init["write_files"])
+    launch = init["runcmd"][1]
+    assert launch[0] == "timeout"
+    assert launch[-3:-1] == ["sh", "-c"]
+    install, bootstrap_command = launch[-1].split(" && ")
+    assert "'azure-storage-blob>=12.24,<13'" in install
+    assert "'azure-identity>=1.19,<2'" in install
+    assert bootstrap_command == "exec python -u /opt/flash/capsule.pyz bootstrap"
 
 
 def test_spot_handle_roundtrips_through_cleanup_identity():
@@ -442,9 +449,14 @@ def test_capsule_publication_pins_revision_and_checksum(monkeypatch):
     from contextlib import contextmanager
 
     @contextmanager
-    def container(token):
-        assert token == "private-capability"
-        yield SimpleNamespace(get_container_properties=lambda: SimpleNamespace(public_access=None))
+    def container(token=None):
+        assert token in (None, "private-capability")
+        # A worker SAS cannot read container properties.
+        yield (
+            SimpleNamespace(get_container_properties=lambda: SimpleNamespace(public_access=None))
+            if token is None
+            else SimpleNamespace()
+        )
 
     monkeypatch.setattr(bootstrap, "open_container", container)
     monkeypatch.setattr(
@@ -555,3 +567,24 @@ def test_orphan_sweep_compares_original_run_ids(monkeypatch, run_id, callbacks):
         (lambda: known) if callbacks else known,
     )
     assert result == deleted == ["terminal"]
+
+
+def test_capsule_rejects_public_container_before_using_worker_sas(monkeypatch):
+    from contextlib import contextmanager
+
+    from flash.providers.nebius import bootstrap
+
+    @contextmanager
+    def container(token=None):
+        assert token is None, "must reject public storage before using the worker capability"
+        yield SimpleNamespace(
+            get_container_properties=lambda: SimpleNamespace(public_access="blob")
+        )
+
+    monkeypatch.setattr(bootstrap, "open_container", container)
+    with pytest.raises(ValueError, match="must be private"):
+        bootstrap.publish_capsule(
+            {"hf_repo": "test/artifacts", "env": {"FLASH_ARTIFACT_BLOB_SAS_URL": "worker-sas"}},
+            b"capsule",
+            deadline_at=time.time() + 60,
+        )

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/test_qwen9b_base.py

@@ -51,8 +51,8 @@ def test_pretrained_serving_uses_its_own_weights_tokenizer_and_revision():
     assert profile.modal_gpu_request == "H100!"
     assert profile.max_model_len == 262144
     assert profile.max_num_seqs == get_model(BASE).serving.max_num_seqs == 32
-    # Hosted FP8 serving uses L40S; the separate customer-owned BF16 profile stays on H100.
-    assert gpu_for(BASE) == gpu_for(POST_TRAINED) == "L40S"
+    # Hosted FP8 serving uses H100; the separate customer-owned BF16 profile stays on H100.
+    assert gpu_for(BASE) == gpu_for(POST_TRAINED) == "H100"
     assert overrides["max_num_seqs"] == 64
     assert "max_num_batched_tokens" not in overrides
     assert tool_parser_for(BASE) == "qwen3_coder"

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/tests/test_serve_profiles.py

@@ -176,8 +176,8 @@ def test_profile_matches_the_catalog_serving_capacity(model_id: str) -> None:
         == {
             "Qwen/Qwen3.5-0.8B": "L4",
             "Qwen/Qwen3.5-4B": "L40S",
-            "Qwen/Qwen3.5-9B": "L40S",
-            "Qwen/Qwen3.5-9B-Base": "L40S",
+            "Qwen/Qwen3.5-9B": "H100",
+            "Qwen/Qwen3.5-9B-Base": "H100",
             "Qwen/Qwen3.8-27B": "H100",
             "Qwen/Qwen3.6-35B-A3B": "H200",
         }[model_id]

{freesolo_flash_dev-1.4.22 → freesolo_flash_dev-1.4.23}/uv.lock

@@ -1297,7 +1297,7 @@ wheels = [
 
 [[package]]
 name = "freesolo-flash"
-version = "1.4.22"
+version = "1.4.23"
 source = { editable = "." }
 
 [package.optional-dependencies]