kalavai-client 0.6.22__py3-none-any.whl → 0.7.0__py3-none-any.whl

kalavai_client/__init__.py

@@ -1,2 +1,2 @@
 
-__version__ = "0.6.22"
+__version__ = "0.7.0"

kalavai_client/assets/apps.yaml

@@ -2,6 +2,13 @@ helmDefaults:
   timeout: 1200
 
 repositories:
+  # amd gpu operator: https://github.com/ROCm/gpu-operator
+  - name: rocm
+    url: https://rocm.github.io/gpu-operator
+  # required by rocm
+  - name: jetstacks
+    url: https://charts.jetstack.io
+  ###
   - name: kuberay
     url: https://ray-project.github.io/kuberay-helm/
   - name: nvidia
@@ -24,6 +31,20 @@ repositories:
     url: https://project-hami.github.io/HAMi
 
 releases:
+  - name: certificates
+    namespace: kalavai
+    chart: jetstacks/cert-manager
+    version: v1.15.1
+    installed: true
+    set:
+    - name: crds.enabled
+      value: true
+  - name: rocm
+    needs: 
+    - kalavai/certificates
+    namespace: kalavai
+    chart: rocm/gpu-operator-charts
+    installed: true
   - name: datashim
     namespace: dlf
     chart: kalavai/kalavai-datashim
@@ -156,7 +177,7 @@ releases:
     - name: replicas
       value: 1
     - name: image_tag
-      value: "v2025.07.34"
+      value: "v2025.08.3" #"v2025.07.34"
     - name: deployment.in_cluster
       value: "True"
     - name: deployment.kalavai_username_key
@@ -186,28 +207,7 @@ releases:
   - name: nvidia-gpu-operator
     namespace: kalavai
     chart: kalavai/gpu
-    installed: false    
-  # - name: hami-vgpu
-  #   namespace: kalavai
-  #   chart: kalavai/hami
-  #   installed: true
-  #   set:
-  #   - name: resourceCores
-  #     value: "nvidia.com/gpucores"
-  #   - name: devicePlugin.runtimeClassName
-  #     value: "nvidia"
-  #   - name: scheduler.defaultSchedulerPolicy.nodeSchedulerPolicy
-  #     value: "binpack"
-  #   - name: scheduler.defaultSchedulerPolicy.gpuSchedulerPolicy
-  #     value: "binpack"
-  #   - name: scheduler.defaultCores
-  #     value: "100"
-  #   - name: scheduler.kubeScheduler.imageTag
-  #     value: v1.31.1
-  #   - name: devicePlugin.deviceMemoryScaling
-  #     value: "1"
-  #   - name: devicePlugin.deviceSplitCount
-  #     value: "1"
+    installed: false
   - name: hami-vgpu
     namespace: kalavai
     chart: hami-charts/hami

kalavai_client/cli.py

@@ -1166,38 +1166,47 @@ def job__delete(name, *others, force_namespace: str=None):
 
 
 @arguably.command
-def job__estimate(billion_parameters, *others, precision=32):
-    """Guesstimate of resources needed based on required memory and current resources"""
-    try:
-        CLUSTER.validate_cluster()
-    except Exception as e:
-        console.log(f"[red]Problems with your pool: {str(e)}")
-        return
-    
-    average_vram = 8
-    required_memory = float(billion_parameters) * (precision / 8) / 1.2
-    available_gpus = load_gpu_models()
-    vrams = []
-    for _, gpus in available_gpus:
-        for model in gpus["gpus"]:
-            vrams.extend([int(model["memory"])/1000] * int(gpus["capacity"]) )
-    vrams = sorted(vrams, reverse=False)
+def job__estimate(
+    *others, 
+    model_size: float,
+    precision: str = "fp16",
+    context_window: int = 2048,
+    batch_size: int = 1,
+    num_layers: int = 32, # total layers, or num_key_value_heads, whatever is minimum (impacts KV cache)
+    hidden_dim: int = 4096,
+    overhead_factor: float = 0.15
+):
+    # Bytes per parameter based on precision
+    precision_bytes = {
+        "fp32": 4,
+        "fp16": 2,
+        "int8": 1,
+        "int4": 0.5
+    }
 
-    console.log(f"There are {len(vrams)} GPUs available ({sum(vrams)}GBs)")
-    console.log(f"A [yellow]{billion_parameters}B[white] model requires [yellow]~{required_memory:.2f}GB vRAM[white] at {precision}bits precision")
+    if precision not in precision_bytes:
+        raise ValueError(f"Unsupported precision: {precision}. Choose from {list(precision_bytes.keys())}")
 
-    if sum(vrams) < required_memory:
-        console.log("Current capacity is insufficient to host the model, but it can be scheduled for when it is!")
-        console.log(f"Average devices have {average_vram}GB vRAM, use {math.ceil(required_memory/(average_vram))} GPU workers")
-    else:
-        current_vram = 0
-        n_devices = 0
-        for mem in vrams:
-            current_vram += mem
-            n_devices += 1
-            if current_vram > required_memory:
-                break
-        console.log(f"Looking at current capacity, use [green]{n_devices} GPU workers[white] for a total [green]{current_vram:.2f} GB vRAM")  
+    # Model parameters
+    total_params = model_size * 1e9
+    model_weights_vram = total_params * precision_bytes[precision] / 1e9  # GB
+
+    # KV Cache memory
+    # Approximation: KV cache = batch × layers × hidden_dim × 2 (K/V) × context × bytes
+    kv_cache_vram = (
+        batch_size * num_layers * hidden_dim * 2 * context_window * precision_bytes[precision]
+    ) / 1e9  # GB
+
+    # Total VRAM including overhead
+    total_vram = model_weights_vram + kv_cache_vram
+    total_vram *= (1 + overhead_factor)
+
+    result = {
+        "model_weights_vram_gb": round(model_weights_vram, 2),
+        "kv_cache_vram_gb": round(kv_cache_vram, 2),
+        "estimated_total_vram_gb": round(total_vram, 2)
+    }
+    console.log(f"[green]{result}")
 
 @arguably.command
 def job__status(name, *others):

{kalavai_client-0.6.22.dist-info → kalavai_client-0.7.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: kalavai-client
-Version: 0.6.22
+Version: 0.7.0
 Summary: Client app for kalavai platform
 License: Apache-2.0
 Keywords: LLM,platform

{kalavai_client-0.6.22.dist-info → kalavai_client-0.7.0.dist-info}/RECORD

@@ -1,7 +1,7 @@
-kalavai_client/__init__.py,sha256=v2m8n1AiRAfLYkbEyWtMgln2mUtzHOjwrvGiSwSCHCg,23
+kalavai_client/__init__.py,sha256=vNC3xHXqvZzke4pHwX0fLN2HjCBaYwJV_rhNXqxKX6I,22
 kalavai_client/__main__.py,sha256=WQUfxvRsBJH5gsCJg8pLz95QnZIj7Ol8psTO77m0QE0,73
 kalavai_client/assets/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-kalavai_client/assets/apps.yaml,sha256=HT1Yl_tPp5ysjn9TRhL7PlP1w67QOYZGs0ge-bQlvF4,6891
+kalavai_client/assets/apps.yaml,sha256=ylXUApravPGFRIFitx60Bx7EqQi8K6xMKcIb51pDevI,6730
 kalavai_client/assets/apps_values.yaml,sha256=LeSNd3PwkIx0wkTIlEk2KNz3Yy4sXSaHALQEkopdhKE,2165
 kalavai_client/assets/docker-compose-gui.yaml,sha256=OAVO0ohaCpDB9FGeih0yAbVNwUfDtaCzssZ25uiuJyA,787
 kalavai_client/assets/docker-compose-template.yaml,sha256=vW7GhOl_PaUodehJk8qajOlE0deZXrPc7qizg5SeYyc,1859
@@ -13,13 +13,13 @@ kalavai_client/assets/user_workspace_values.yaml,sha256=G0HOzQUxrDMCwuW9kbWUZaKM
 kalavai_client/auth.py,sha256=EB3PMvKUn5_KAQkezkEHEt-OMZXyfkZguIQlUFkEHcA,3243
 kalavai_client/bridge_api.py,sha256=qiN0jleaooj2vYVYxHrG_nfdgY2rKpZyvFE4uz6hkoA,27088
 kalavai_client/bridge_models.py,sha256=mUh67hzhudqGxJEFHx2KGrf-Sjrt_CwkzLno8Xsm7hk,3043
-kalavai_client/cli.py,sha256=ZRNOv1oUvU7Freu47PotrwqJRrBMSFgmoCHg620UdZM,48146
+kalavai_client/cli.py,sha256=QihLex32dFmneV8-JhMA7J1BXqOp1mI-IJH37O6-EzE,48007
 kalavai_client/cluster.py,sha256=Z2PIXbZuSAv9xmw-MyZP1M41BpVMpirLzG51bqGA-zc,13548
 kalavai_client/core.py,sha256=JVXSMmYvbNBl9ggVPGNJRryK54doySTrGDj-WhAlkfY,35760
 kalavai_client/env.py,sha256=t6dfjg5GY6lbprbmlr9dVOP_KouPwdN94wnDL5zCgIM,2902
 kalavai_client/utils.py,sha256=1mz-dzoJhZ9GJKU7jiGYBC1tP37SXHvxToMqqEir8R0,13438
-kalavai_client-0.6.22.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-kalavai_client-0.6.22.dist-info/METADATA,sha256=uquNcK5cPCexquqwajy7XvqBC1LEE2USs677qtcFqMg,12776
-kalavai_client-0.6.22.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-kalavai_client-0.6.22.dist-info/entry_points.txt,sha256=9T6D45gxwzfVbglMm1r6XPdXuuZdHfy_7fCeu2jUphc,50
-kalavai_client-0.6.22.dist-info/RECORD,,
+kalavai_client-0.7.0.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+kalavai_client-0.7.0.dist-info/METADATA,sha256=4cWmWAY7QZHbrVVsTIfI_d7yp3thrAXagOVBr1hJ9fk,12775
+kalavai_client-0.7.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+kalavai_client-0.7.0.dist-info/entry_points.txt,sha256=9T6D45gxwzfVbglMm1r6XPdXuuZdHfy_7fCeu2jUphc,50
+kalavai_client-0.7.0.dist-info/RECORD,,